Loading...
1/*
2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <bcrl@kvack.org>
4 *
5 * Implements an efficient asynchronous io interface.
6 *
7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved.
8 *
9 * See ../COPYING for licensing terms.
10 */
11#include <linux/kernel.h>
12#include <linux/init.h>
13#include <linux/errno.h>
14#include <linux/time.h>
15#include <linux/aio_abi.h>
16#include <linux/module.h>
17#include <linux/syscalls.h>
18#include <linux/backing-dev.h>
19#include <linux/uio.h>
20
21#define DEBUG 0
22
23#include <linux/sched.h>
24#include <linux/fs.h>
25#include <linux/file.h>
26#include <linux/mm.h>
27#include <linux/mman.h>
28#include <linux/mmu_context.h>
29#include <linux/slab.h>
30#include <linux/timer.h>
31#include <linux/aio.h>
32#include <linux/highmem.h>
33#include <linux/workqueue.h>
34#include <linux/security.h>
35#include <linux/eventfd.h>
36#include <linux/blkdev.h>
37#include <linux/compat.h>
38
39#include <asm/kmap_types.h>
40#include <asm/uaccess.h>
41
42#if DEBUG > 1
43#define dprintk printk
44#else
45#define dprintk(x...) do { ; } while (0)
46#endif
47
48/*------ sysctl variables----*/
49static DEFINE_SPINLOCK(aio_nr_lock);
50unsigned long aio_nr; /* current system wide number of aio requests */
51unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
52/*----end sysctl variables---*/
53
54static struct kmem_cache *kiocb_cachep;
55static struct kmem_cache *kioctx_cachep;
56
57static struct workqueue_struct *aio_wq;
58
59/* Used for rare fput completion. */
60static void aio_fput_routine(struct work_struct *);
61static DECLARE_WORK(fput_work, aio_fput_routine);
62
63static DEFINE_SPINLOCK(fput_lock);
64static LIST_HEAD(fput_head);
65
66static void aio_kick_handler(struct work_struct *);
67static void aio_queue_work(struct kioctx *);
68
69/* aio_setup
70 * Creates the slab caches used by the aio routines, panic on
71 * failure as this is done early during the boot sequence.
72 */
73static int __init aio_setup(void)
74{
75 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
76 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
77
78 aio_wq = alloc_workqueue("aio", 0, 1); /* used to limit concurrency */
79 BUG_ON(!aio_wq);
80
81 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
82
83 return 0;
84}
85__initcall(aio_setup);
86
87static void aio_free_ring(struct kioctx *ctx)
88{
89 struct aio_ring_info *info = &ctx->ring_info;
90 long i;
91
92 for (i=0; i<info->nr_pages; i++)
93 put_page(info->ring_pages[i]);
94
95 if (info->mmap_size) {
96 down_write(&ctx->mm->mmap_sem);
97 do_munmap(ctx->mm, info->mmap_base, info->mmap_size);
98 up_write(&ctx->mm->mmap_sem);
99 }
100
101 if (info->ring_pages && info->ring_pages != info->internal_pages)
102 kfree(info->ring_pages);
103 info->ring_pages = NULL;
104 info->nr = 0;
105}
106
107static int aio_setup_ring(struct kioctx *ctx)
108{
109 struct aio_ring *ring;
110 struct aio_ring_info *info = &ctx->ring_info;
111 unsigned nr_events = ctx->max_reqs;
112 unsigned long size;
113 int nr_pages;
114
115 /* Compensate for the ring buffer's head/tail overlap entry */
116 nr_events += 2; /* 1 is required, 2 for good luck */
117
118 size = sizeof(struct aio_ring);
119 size += sizeof(struct io_event) * nr_events;
120 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
121
122 if (nr_pages < 0)
123 return -EINVAL;
124
125 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
126
127 info->nr = 0;
128 info->ring_pages = info->internal_pages;
129 if (nr_pages > AIO_RING_PAGES) {
130 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
131 if (!info->ring_pages)
132 return -ENOMEM;
133 }
134
135 info->mmap_size = nr_pages * PAGE_SIZE;
136 dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
137 down_write(&ctx->mm->mmap_sem);
138 info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
139 PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE,
140 0);
141 if (IS_ERR((void *)info->mmap_base)) {
142 up_write(&ctx->mm->mmap_sem);
143 info->mmap_size = 0;
144 aio_free_ring(ctx);
145 return -EAGAIN;
146 }
147
148 dprintk("mmap address: 0x%08lx\n", info->mmap_base);
149 info->nr_pages = get_user_pages(current, ctx->mm,
150 info->mmap_base, nr_pages,
151 1, 0, info->ring_pages, NULL);
152 up_write(&ctx->mm->mmap_sem);
153
154 if (unlikely(info->nr_pages != nr_pages)) {
155 aio_free_ring(ctx);
156 return -EAGAIN;
157 }
158
159 ctx->user_id = info->mmap_base;
160
161 info->nr = nr_events; /* trusted copy */
162
163 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
164 ring->nr = nr_events; /* user copy */
165 ring->id = ctx->user_id;
166 ring->head = ring->tail = 0;
167 ring->magic = AIO_RING_MAGIC;
168 ring->compat_features = AIO_RING_COMPAT_FEATURES;
169 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
170 ring->header_length = sizeof(struct aio_ring);
171 kunmap_atomic(ring, KM_USER0);
172
173 return 0;
174}
175
176
177/* aio_ring_event: returns a pointer to the event at the given index from
178 * kmap_atomic(, km). Release the pointer with put_aio_ring_event();
179 */
180#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
181#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
182#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
183
184#define aio_ring_event(info, nr, km) ({ \
185 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
186 struct io_event *__event; \
187 __event = kmap_atomic( \
188 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
189 __event += pos % AIO_EVENTS_PER_PAGE; \
190 __event; \
191})
192
193#define put_aio_ring_event(event, km) do { \
194 struct io_event *__event = (event); \
195 (void)__event; \
196 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
197} while(0)
198
199static void ctx_rcu_free(struct rcu_head *head)
200{
201 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
202 unsigned nr_events = ctx->max_reqs;
203
204 kmem_cache_free(kioctx_cachep, ctx);
205
206 if (nr_events) {
207 spin_lock(&aio_nr_lock);
208 BUG_ON(aio_nr - nr_events > aio_nr);
209 aio_nr -= nr_events;
210 spin_unlock(&aio_nr_lock);
211 }
212}
213
214/* __put_ioctx
215 * Called when the last user of an aio context has gone away,
216 * and the struct needs to be freed.
217 */
218static void __put_ioctx(struct kioctx *ctx)
219{
220 BUG_ON(ctx->reqs_active);
221
222 cancel_delayed_work(&ctx->wq);
223 cancel_work_sync(&ctx->wq.work);
224 aio_free_ring(ctx);
225 mmdrop(ctx->mm);
226 ctx->mm = NULL;
227 pr_debug("__put_ioctx: freeing %p\n", ctx);
228 call_rcu(&ctx->rcu_head, ctx_rcu_free);
229}
230
231static inline void get_ioctx(struct kioctx *kioctx)
232{
233 BUG_ON(atomic_read(&kioctx->users) <= 0);
234 atomic_inc(&kioctx->users);
235}
236
237static inline int try_get_ioctx(struct kioctx *kioctx)
238{
239 return atomic_inc_not_zero(&kioctx->users);
240}
241
242static inline void put_ioctx(struct kioctx *kioctx)
243{
244 BUG_ON(atomic_read(&kioctx->users) <= 0);
245 if (unlikely(atomic_dec_and_test(&kioctx->users)))
246 __put_ioctx(kioctx);
247}
248
249/* ioctx_alloc
250 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
251 */
252static struct kioctx *ioctx_alloc(unsigned nr_events)
253{
254 struct mm_struct *mm;
255 struct kioctx *ctx;
256 int did_sync = 0;
257
258 /* Prevent overflows */
259 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
260 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
261 pr_debug("ENOMEM: nr_events too high\n");
262 return ERR_PTR(-EINVAL);
263 }
264
265 if ((unsigned long)nr_events > aio_max_nr)
266 return ERR_PTR(-EAGAIN);
267
268 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
269 if (!ctx)
270 return ERR_PTR(-ENOMEM);
271
272 ctx->max_reqs = nr_events;
273 mm = ctx->mm = current->mm;
274 atomic_inc(&mm->mm_count);
275
276 atomic_set(&ctx->users, 1);
277 spin_lock_init(&ctx->ctx_lock);
278 spin_lock_init(&ctx->ring_info.ring_lock);
279 init_waitqueue_head(&ctx->wait);
280
281 INIT_LIST_HEAD(&ctx->active_reqs);
282 INIT_LIST_HEAD(&ctx->run_list);
283 INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
284
285 if (aio_setup_ring(ctx) < 0)
286 goto out_freectx;
287
288 /* limit the number of system wide aios */
289 do {
290 spin_lock_bh(&aio_nr_lock);
291 if (aio_nr + nr_events > aio_max_nr ||
292 aio_nr + nr_events < aio_nr)
293 ctx->max_reqs = 0;
294 else
295 aio_nr += ctx->max_reqs;
296 spin_unlock_bh(&aio_nr_lock);
297 if (ctx->max_reqs || did_sync)
298 break;
299
300 /* wait for rcu callbacks to have completed before giving up */
301 synchronize_rcu();
302 did_sync = 1;
303 ctx->max_reqs = nr_events;
304 } while (1);
305
306 if (ctx->max_reqs == 0)
307 goto out_cleanup;
308
309 /* now link into global list. */
310 spin_lock(&mm->ioctx_lock);
311 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
312 spin_unlock(&mm->ioctx_lock);
313
314 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
315 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
316 return ctx;
317
318out_cleanup:
319 __put_ioctx(ctx);
320 return ERR_PTR(-EAGAIN);
321
322out_freectx:
323 mmdrop(mm);
324 kmem_cache_free(kioctx_cachep, ctx);
325 ctx = ERR_PTR(-ENOMEM);
326
327 dprintk("aio: error allocating ioctx %p\n", ctx);
328 return ctx;
329}
330
331/* aio_cancel_all
332 * Cancels all outstanding aio requests on an aio context. Used
333 * when the processes owning a context have all exited to encourage
334 * the rapid destruction of the kioctx.
335 */
336static void aio_cancel_all(struct kioctx *ctx)
337{
338 int (*cancel)(struct kiocb *, struct io_event *);
339 struct io_event res;
340 spin_lock_irq(&ctx->ctx_lock);
341 ctx->dead = 1;
342 while (!list_empty(&ctx->active_reqs)) {
343 struct list_head *pos = ctx->active_reqs.next;
344 struct kiocb *iocb = list_kiocb(pos);
345 list_del_init(&iocb->ki_list);
346 cancel = iocb->ki_cancel;
347 kiocbSetCancelled(iocb);
348 if (cancel) {
349 iocb->ki_users++;
350 spin_unlock_irq(&ctx->ctx_lock);
351 cancel(iocb, &res);
352 spin_lock_irq(&ctx->ctx_lock);
353 }
354 }
355 spin_unlock_irq(&ctx->ctx_lock);
356}
357
358static void wait_for_all_aios(struct kioctx *ctx)
359{
360 struct task_struct *tsk = current;
361 DECLARE_WAITQUEUE(wait, tsk);
362
363 spin_lock_irq(&ctx->ctx_lock);
364 if (!ctx->reqs_active)
365 goto out;
366
367 add_wait_queue(&ctx->wait, &wait);
368 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
369 while (ctx->reqs_active) {
370 spin_unlock_irq(&ctx->ctx_lock);
371 io_schedule();
372 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
373 spin_lock_irq(&ctx->ctx_lock);
374 }
375 __set_task_state(tsk, TASK_RUNNING);
376 remove_wait_queue(&ctx->wait, &wait);
377
378out:
379 spin_unlock_irq(&ctx->ctx_lock);
380}
381
382/* wait_on_sync_kiocb:
383 * Waits on the given sync kiocb to complete.
384 */
385ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
386{
387 while (iocb->ki_users) {
388 set_current_state(TASK_UNINTERRUPTIBLE);
389 if (!iocb->ki_users)
390 break;
391 io_schedule();
392 }
393 __set_current_state(TASK_RUNNING);
394 return iocb->ki_user_data;
395}
396EXPORT_SYMBOL(wait_on_sync_kiocb);
397
398/* exit_aio: called when the last user of mm goes away. At this point,
399 * there is no way for any new requests to be submited or any of the
400 * io_* syscalls to be called on the context. However, there may be
401 * outstanding requests which hold references to the context; as they
402 * go away, they will call put_ioctx and release any pinned memory
403 * associated with the request (held via struct page * references).
404 */
405void exit_aio(struct mm_struct *mm)
406{
407 struct kioctx *ctx;
408
409 while (!hlist_empty(&mm->ioctx_list)) {
410 ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
411 hlist_del_rcu(&ctx->list);
412
413 aio_cancel_all(ctx);
414
415 wait_for_all_aios(ctx);
416 /*
417 * Ensure we don't leave the ctx on the aio_wq
418 */
419 cancel_work_sync(&ctx->wq.work);
420
421 if (1 != atomic_read(&ctx->users))
422 printk(KERN_DEBUG
423 "exit_aio:ioctx still alive: %d %d %d\n",
424 atomic_read(&ctx->users), ctx->dead,
425 ctx->reqs_active);
426 put_ioctx(ctx);
427 }
428}
429
430/* aio_get_req
431 * Allocate a slot for an aio request. Increments the users count
432 * of the kioctx so that the kioctx stays around until all requests are
433 * complete. Returns NULL if no requests are free.
434 *
435 * Returns with kiocb->users set to 2. The io submit code path holds
436 * an extra reference while submitting the i/o.
437 * This prevents races between the aio code path referencing the
438 * req (after submitting it) and aio_complete() freeing the req.
439 */
440static struct kiocb *__aio_get_req(struct kioctx *ctx)
441{
442 struct kiocb *req = NULL;
443 struct aio_ring *ring;
444 int okay = 0;
445
446 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
447 if (unlikely(!req))
448 return NULL;
449
450 req->ki_flags = 0;
451 req->ki_users = 2;
452 req->ki_key = 0;
453 req->ki_ctx = ctx;
454 req->ki_cancel = NULL;
455 req->ki_retry = NULL;
456 req->ki_dtor = NULL;
457 req->private = NULL;
458 req->ki_iovec = NULL;
459 INIT_LIST_HEAD(&req->ki_run_list);
460 req->ki_eventfd = NULL;
461
462 /* Check if the completion queue has enough free space to
463 * accept an event from this io.
464 */
465 spin_lock_irq(&ctx->ctx_lock);
466 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
467 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
468 list_add(&req->ki_list, &ctx->active_reqs);
469 ctx->reqs_active++;
470 okay = 1;
471 }
472 kunmap_atomic(ring, KM_USER0);
473 spin_unlock_irq(&ctx->ctx_lock);
474
475 if (!okay) {
476 kmem_cache_free(kiocb_cachep, req);
477 req = NULL;
478 }
479
480 return req;
481}
482
483static inline struct kiocb *aio_get_req(struct kioctx *ctx)
484{
485 struct kiocb *req;
486 /* Handle a potential starvation case -- should be exceedingly rare as
487 * requests will be stuck on fput_head only if the aio_fput_routine is
488 * delayed and the requests were the last user of the struct file.
489 */
490 req = __aio_get_req(ctx);
491 if (unlikely(NULL == req)) {
492 aio_fput_routine(NULL);
493 req = __aio_get_req(ctx);
494 }
495 return req;
496}
497
498static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
499{
500 assert_spin_locked(&ctx->ctx_lock);
501
502 if (req->ki_eventfd != NULL)
503 eventfd_ctx_put(req->ki_eventfd);
504 if (req->ki_dtor)
505 req->ki_dtor(req);
506 if (req->ki_iovec != &req->ki_inline_vec)
507 kfree(req->ki_iovec);
508 kmem_cache_free(kiocb_cachep, req);
509 ctx->reqs_active--;
510
511 if (unlikely(!ctx->reqs_active && ctx->dead))
512 wake_up_all(&ctx->wait);
513}
514
515static void aio_fput_routine(struct work_struct *data)
516{
517 spin_lock_irq(&fput_lock);
518 while (likely(!list_empty(&fput_head))) {
519 struct kiocb *req = list_kiocb(fput_head.next);
520 struct kioctx *ctx = req->ki_ctx;
521
522 list_del(&req->ki_list);
523 spin_unlock_irq(&fput_lock);
524
525 /* Complete the fput(s) */
526 if (req->ki_filp != NULL)
527 fput(req->ki_filp);
528
529 /* Link the iocb into the context's free list */
530 spin_lock_irq(&ctx->ctx_lock);
531 really_put_req(ctx, req);
532 spin_unlock_irq(&ctx->ctx_lock);
533
534 put_ioctx(ctx);
535 spin_lock_irq(&fput_lock);
536 }
537 spin_unlock_irq(&fput_lock);
538}
539
540/* __aio_put_req
541 * Returns true if this put was the last user of the request.
542 */
543static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
544{
545 dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
546 req, atomic_long_read(&req->ki_filp->f_count));
547
548 assert_spin_locked(&ctx->ctx_lock);
549
550 req->ki_users--;
551 BUG_ON(req->ki_users < 0);
552 if (likely(req->ki_users))
553 return 0;
554 list_del(&req->ki_list); /* remove from active_reqs */
555 req->ki_cancel = NULL;
556 req->ki_retry = NULL;
557
558 /*
559 * Try to optimize the aio and eventfd file* puts, by avoiding to
560 * schedule work in case it is not final fput() time. In normal cases,
561 * we would not be holding the last reference to the file*, so
562 * this function will be executed w/out any aio kthread wakeup.
563 */
564 if (unlikely(!fput_atomic(req->ki_filp))) {
565 get_ioctx(ctx);
566 spin_lock(&fput_lock);
567 list_add(&req->ki_list, &fput_head);
568 spin_unlock(&fput_lock);
569 schedule_work(&fput_work);
570 } else {
571 req->ki_filp = NULL;
572 really_put_req(ctx, req);
573 }
574 return 1;
575}
576
577/* aio_put_req
578 * Returns true if this put was the last user of the kiocb,
579 * false if the request is still in use.
580 */
581int aio_put_req(struct kiocb *req)
582{
583 struct kioctx *ctx = req->ki_ctx;
584 int ret;
585 spin_lock_irq(&ctx->ctx_lock);
586 ret = __aio_put_req(ctx, req);
587 spin_unlock_irq(&ctx->ctx_lock);
588 return ret;
589}
590EXPORT_SYMBOL(aio_put_req);
591
592static struct kioctx *lookup_ioctx(unsigned long ctx_id)
593{
594 struct mm_struct *mm = current->mm;
595 struct kioctx *ctx, *ret = NULL;
596 struct hlist_node *n;
597
598 rcu_read_lock();
599
600 hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) {
601 /*
602 * RCU protects us against accessing freed memory but
603 * we have to be careful not to get a reference when the
604 * reference count already dropped to 0 (ctx->dead test
605 * is unreliable because of races).
606 */
607 if (ctx->user_id == ctx_id && !ctx->dead && try_get_ioctx(ctx)){
608 ret = ctx;
609 break;
610 }
611 }
612
613 rcu_read_unlock();
614 return ret;
615}
616
617/*
618 * Queue up a kiocb to be retried. Assumes that the kiocb
619 * has already been marked as kicked, and places it on
620 * the retry run list for the corresponding ioctx, if it
621 * isn't already queued. Returns 1 if it actually queued
622 * the kiocb (to tell the caller to activate the work
623 * queue to process it), or 0, if it found that it was
624 * already queued.
625 */
626static inline int __queue_kicked_iocb(struct kiocb *iocb)
627{
628 struct kioctx *ctx = iocb->ki_ctx;
629
630 assert_spin_locked(&ctx->ctx_lock);
631
632 if (list_empty(&iocb->ki_run_list)) {
633 list_add_tail(&iocb->ki_run_list,
634 &ctx->run_list);
635 return 1;
636 }
637 return 0;
638}
639
640/* aio_run_iocb
641 * This is the core aio execution routine. It is
642 * invoked both for initial i/o submission and
643 * subsequent retries via the aio_kick_handler.
644 * Expects to be invoked with iocb->ki_ctx->lock
645 * already held. The lock is released and reacquired
646 * as needed during processing.
647 *
648 * Calls the iocb retry method (already setup for the
649 * iocb on initial submission) for operation specific
650 * handling, but takes care of most of common retry
651 * execution details for a given iocb. The retry method
652 * needs to be non-blocking as far as possible, to avoid
653 * holding up other iocbs waiting to be serviced by the
654 * retry kernel thread.
655 *
656 * The trickier parts in this code have to do with
657 * ensuring that only one retry instance is in progress
658 * for a given iocb at any time. Providing that guarantee
659 * simplifies the coding of individual aio operations as
660 * it avoids various potential races.
661 */
662static ssize_t aio_run_iocb(struct kiocb *iocb)
663{
664 struct kioctx *ctx = iocb->ki_ctx;
665 ssize_t (*retry)(struct kiocb *);
666 ssize_t ret;
667
668 if (!(retry = iocb->ki_retry)) {
669 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
670 return 0;
671 }
672
673 /*
674 * We don't want the next retry iteration for this
675 * operation to start until this one has returned and
676 * updated the iocb state. However, wait_queue functions
677 * can trigger a kick_iocb from interrupt context in the
678 * meantime, indicating that data is available for the next
679 * iteration. We want to remember that and enable the
680 * next retry iteration _after_ we are through with
681 * this one.
682 *
683 * So, in order to be able to register a "kick", but
684 * prevent it from being queued now, we clear the kick
685 * flag, but make the kick code *think* that the iocb is
686 * still on the run list until we are actually done.
687 * When we are done with this iteration, we check if
688 * the iocb was kicked in the meantime and if so, queue
689 * it up afresh.
690 */
691
692 kiocbClearKicked(iocb);
693
694 /*
695 * This is so that aio_complete knows it doesn't need to
696 * pull the iocb off the run list (We can't just call
697 * INIT_LIST_HEAD because we don't want a kick_iocb to
698 * queue this on the run list yet)
699 */
700 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
701 spin_unlock_irq(&ctx->ctx_lock);
702
703 /* Quit retrying if the i/o has been cancelled */
704 if (kiocbIsCancelled(iocb)) {
705 ret = -EINTR;
706 aio_complete(iocb, ret, 0);
707 /* must not access the iocb after this */
708 goto out;
709 }
710
711 /*
712 * Now we are all set to call the retry method in async
713 * context.
714 */
715 ret = retry(iocb);
716
717 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
718 /*
719 * There's no easy way to restart the syscall since other AIO's
720 * may be already running. Just fail this IO with EINTR.
721 */
722 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
723 ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
724 ret = -EINTR;
725 aio_complete(iocb, ret, 0);
726 }
727out:
728 spin_lock_irq(&ctx->ctx_lock);
729
730 if (-EIOCBRETRY == ret) {
731 /*
732 * OK, now that we are done with this iteration
733 * and know that there is more left to go,
734 * this is where we let go so that a subsequent
735 * "kick" can start the next iteration
736 */
737
738 /* will make __queue_kicked_iocb succeed from here on */
739 INIT_LIST_HEAD(&iocb->ki_run_list);
740 /* we must queue the next iteration ourselves, if it
741 * has already been kicked */
742 if (kiocbIsKicked(iocb)) {
743 __queue_kicked_iocb(iocb);
744
745 /*
746 * __queue_kicked_iocb will always return 1 here, because
747 * iocb->ki_run_list is empty at this point so it should
748 * be safe to unconditionally queue the context into the
749 * work queue.
750 */
751 aio_queue_work(ctx);
752 }
753 }
754 return ret;
755}
756
757/*
758 * __aio_run_iocbs:
759 * Process all pending retries queued on the ioctx
760 * run list.
761 * Assumes it is operating within the aio issuer's mm
762 * context.
763 */
764static int __aio_run_iocbs(struct kioctx *ctx)
765{
766 struct kiocb *iocb;
767 struct list_head run_list;
768
769 assert_spin_locked(&ctx->ctx_lock);
770
771 list_replace_init(&ctx->run_list, &run_list);
772 while (!list_empty(&run_list)) {
773 iocb = list_entry(run_list.next, struct kiocb,
774 ki_run_list);
775 list_del(&iocb->ki_run_list);
776 /*
777 * Hold an extra reference while retrying i/o.
778 */
779 iocb->ki_users++; /* grab extra reference */
780 aio_run_iocb(iocb);
781 __aio_put_req(ctx, iocb);
782 }
783 if (!list_empty(&ctx->run_list))
784 return 1;
785 return 0;
786}
787
788static void aio_queue_work(struct kioctx * ctx)
789{
790 unsigned long timeout;
791 /*
792 * if someone is waiting, get the work started right
793 * away, otherwise, use a longer delay
794 */
795 smp_mb();
796 if (waitqueue_active(&ctx->wait))
797 timeout = 1;
798 else
799 timeout = HZ/10;
800 queue_delayed_work(aio_wq, &ctx->wq, timeout);
801}
802
803/*
804 * aio_run_all_iocbs:
805 * Process all pending retries queued on the ioctx
806 * run list, and keep running them until the list
807 * stays empty.
808 * Assumes it is operating within the aio issuer's mm context.
809 */
810static inline void aio_run_all_iocbs(struct kioctx *ctx)
811{
812 spin_lock_irq(&ctx->ctx_lock);
813 while (__aio_run_iocbs(ctx))
814 ;
815 spin_unlock_irq(&ctx->ctx_lock);
816}
817
818/*
819 * aio_kick_handler:
820 * Work queue handler triggered to process pending
821 * retries on an ioctx. Takes on the aio issuer's
822 * mm context before running the iocbs, so that
823 * copy_xxx_user operates on the issuer's address
824 * space.
825 * Run on aiod's context.
826 */
827static void aio_kick_handler(struct work_struct *work)
828{
829 struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
830 mm_segment_t oldfs = get_fs();
831 struct mm_struct *mm;
832 int requeue;
833
834 set_fs(USER_DS);
835 use_mm(ctx->mm);
836 spin_lock_irq(&ctx->ctx_lock);
837 requeue =__aio_run_iocbs(ctx);
838 mm = ctx->mm;
839 spin_unlock_irq(&ctx->ctx_lock);
840 unuse_mm(mm);
841 set_fs(oldfs);
842 /*
843 * we're in a worker thread already, don't use queue_delayed_work,
844 */
845 if (requeue)
846 queue_delayed_work(aio_wq, &ctx->wq, 0);
847}
848
849
850/*
851 * Called by kick_iocb to queue the kiocb for retry
852 * and if required activate the aio work queue to process
853 * it
854 */
855static void try_queue_kicked_iocb(struct kiocb *iocb)
856{
857 struct kioctx *ctx = iocb->ki_ctx;
858 unsigned long flags;
859 int run = 0;
860
861 spin_lock_irqsave(&ctx->ctx_lock, flags);
862 /* set this inside the lock so that we can't race with aio_run_iocb()
863 * testing it and putting the iocb on the run list under the lock */
864 if (!kiocbTryKick(iocb))
865 run = __queue_kicked_iocb(iocb);
866 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
867 if (run)
868 aio_queue_work(ctx);
869}
870
871/*
872 * kick_iocb:
873 * Called typically from a wait queue callback context
874 * to trigger a retry of the iocb.
875 * The retry is usually executed by aio workqueue
876 * threads (See aio_kick_handler).
877 */
878void kick_iocb(struct kiocb *iocb)
879{
880 /* sync iocbs are easy: they can only ever be executing from a
881 * single context. */
882 if (is_sync_kiocb(iocb)) {
883 kiocbSetKicked(iocb);
884 wake_up_process(iocb->ki_obj.tsk);
885 return;
886 }
887
888 try_queue_kicked_iocb(iocb);
889}
890EXPORT_SYMBOL(kick_iocb);
891
892/* aio_complete
893 * Called when the io request on the given iocb is complete.
894 * Returns true if this is the last user of the request. The
895 * only other user of the request can be the cancellation code.
896 */
897int aio_complete(struct kiocb *iocb, long res, long res2)
898{
899 struct kioctx *ctx = iocb->ki_ctx;
900 struct aio_ring_info *info;
901 struct aio_ring *ring;
902 struct io_event *event;
903 unsigned long flags;
904 unsigned long tail;
905 int ret;
906
907 /*
908 * Special case handling for sync iocbs:
909 * - events go directly into the iocb for fast handling
910 * - the sync task with the iocb in its stack holds the single iocb
911 * ref, no other paths have a way to get another ref
912 * - the sync task helpfully left a reference to itself in the iocb
913 */
914 if (is_sync_kiocb(iocb)) {
915 BUG_ON(iocb->ki_users != 1);
916 iocb->ki_user_data = res;
917 iocb->ki_users = 0;
918 wake_up_process(iocb->ki_obj.tsk);
919 return 1;
920 }
921
922 info = &ctx->ring_info;
923
924 /* add a completion event to the ring buffer.
925 * must be done holding ctx->ctx_lock to prevent
926 * other code from messing with the tail
927 * pointer since we might be called from irq
928 * context.
929 */
930 spin_lock_irqsave(&ctx->ctx_lock, flags);
931
932 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
933 list_del_init(&iocb->ki_run_list);
934
935 /*
936 * cancelled requests don't get events, userland was given one
937 * when the event got cancelled.
938 */
939 if (kiocbIsCancelled(iocb))
940 goto put_rq;
941
942 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
943
944 tail = info->tail;
945 event = aio_ring_event(info, tail, KM_IRQ0);
946 if (++tail >= info->nr)
947 tail = 0;
948
949 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
950 event->data = iocb->ki_user_data;
951 event->res = res;
952 event->res2 = res2;
953
954 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
955 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
956 res, res2);
957
958 /* after flagging the request as done, we
959 * must never even look at it again
960 */
961 smp_wmb(); /* make event visible before updating tail */
962
963 info->tail = tail;
964 ring->tail = tail;
965
966 put_aio_ring_event(event, KM_IRQ0);
967 kunmap_atomic(ring, KM_IRQ1);
968
969 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
970
971 /*
972 * Check if the user asked us to deliver the result through an
973 * eventfd. The eventfd_signal() function is safe to be called
974 * from IRQ context.
975 */
976 if (iocb->ki_eventfd != NULL)
977 eventfd_signal(iocb->ki_eventfd, 1);
978
979put_rq:
980 /* everything turned out well, dispose of the aiocb. */
981 ret = __aio_put_req(ctx, iocb);
982
983 /*
984 * We have to order our ring_info tail store above and test
985 * of the wait list below outside the wait lock. This is
986 * like in wake_up_bit() where clearing a bit has to be
987 * ordered with the unlocked test.
988 */
989 smp_mb();
990
991 if (waitqueue_active(&ctx->wait))
992 wake_up(&ctx->wait);
993
994 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
995 return ret;
996}
997EXPORT_SYMBOL(aio_complete);
998
999/* aio_read_evt
1000 * Pull an event off of the ioctx's event ring. Returns the number of
1001 * events fetched (0 or 1 ;-)
1002 * FIXME: make this use cmpxchg.
1003 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1004 */
1005static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1006{
1007 struct aio_ring_info *info = &ioctx->ring_info;
1008 struct aio_ring *ring;
1009 unsigned long head;
1010 int ret = 0;
1011
1012 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1013 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1014 (unsigned long)ring->head, (unsigned long)ring->tail,
1015 (unsigned long)ring->nr);
1016
1017 if (ring->head == ring->tail)
1018 goto out;
1019
1020 spin_lock(&info->ring_lock);
1021
1022 head = ring->head % info->nr;
1023 if (head != ring->tail) {
1024 struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1025 *ent = *evp;
1026 head = (head + 1) % info->nr;
1027 smp_mb(); /* finish reading the event before updatng the head */
1028 ring->head = head;
1029 ret = 1;
1030 put_aio_ring_event(evp, KM_USER1);
1031 }
1032 spin_unlock(&info->ring_lock);
1033
1034out:
1035 kunmap_atomic(ring, KM_USER0);
1036 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1037 (unsigned long)ring->head, (unsigned long)ring->tail);
1038 return ret;
1039}
1040
1041struct aio_timeout {
1042 struct timer_list timer;
1043 int timed_out;
1044 struct task_struct *p;
1045};
1046
1047static void timeout_func(unsigned long data)
1048{
1049 struct aio_timeout *to = (struct aio_timeout *)data;
1050
1051 to->timed_out = 1;
1052 wake_up_process(to->p);
1053}
1054
1055static inline void init_timeout(struct aio_timeout *to)
1056{
1057 setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
1058 to->timed_out = 0;
1059 to->p = current;
1060}
1061
1062static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1063 const struct timespec *ts)
1064{
1065 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1066 if (time_after(to->timer.expires, jiffies))
1067 add_timer(&to->timer);
1068 else
1069 to->timed_out = 1;
1070}
1071
1072static inline void clear_timeout(struct aio_timeout *to)
1073{
1074 del_singleshot_timer_sync(&to->timer);
1075}
1076
1077static int read_events(struct kioctx *ctx,
1078 long min_nr, long nr,
1079 struct io_event __user *event,
1080 struct timespec __user *timeout)
1081{
1082 long start_jiffies = jiffies;
1083 struct task_struct *tsk = current;
1084 DECLARE_WAITQUEUE(wait, tsk);
1085 int ret;
1086 int i = 0;
1087 struct io_event ent;
1088 struct aio_timeout to;
1089 int retry = 0;
1090
1091 /* needed to zero any padding within an entry (there shouldn't be
1092 * any, but C is fun!
1093 */
1094 memset(&ent, 0, sizeof(ent));
1095retry:
1096 ret = 0;
1097 while (likely(i < nr)) {
1098 ret = aio_read_evt(ctx, &ent);
1099 if (unlikely(ret <= 0))
1100 break;
1101
1102 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1103 ent.data, ent.obj, ent.res, ent.res2);
1104
1105 /* Could we split the check in two? */
1106 ret = -EFAULT;
1107 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1108 dprintk("aio: lost an event due to EFAULT.\n");
1109 break;
1110 }
1111 ret = 0;
1112
1113 /* Good, event copied to userland, update counts. */
1114 event ++;
1115 i ++;
1116 }
1117
1118 if (min_nr <= i)
1119 return i;
1120 if (ret)
1121 return ret;
1122
1123 /* End fast path */
1124
1125 /* racey check, but it gets redone */
1126 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1127 retry = 1;
1128 aio_run_all_iocbs(ctx);
1129 goto retry;
1130 }
1131
1132 init_timeout(&to);
1133 if (timeout) {
1134 struct timespec ts;
1135 ret = -EFAULT;
1136 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1137 goto out;
1138
1139 set_timeout(start_jiffies, &to, &ts);
1140 }
1141
1142 while (likely(i < nr)) {
1143 add_wait_queue_exclusive(&ctx->wait, &wait);
1144 do {
1145 set_task_state(tsk, TASK_INTERRUPTIBLE);
1146 ret = aio_read_evt(ctx, &ent);
1147 if (ret)
1148 break;
1149 if (min_nr <= i)
1150 break;
1151 if (unlikely(ctx->dead)) {
1152 ret = -EINVAL;
1153 break;
1154 }
1155 if (to.timed_out) /* Only check after read evt */
1156 break;
1157 /* Try to only show up in io wait if there are ops
1158 * in flight */
1159 if (ctx->reqs_active)
1160 io_schedule();
1161 else
1162 schedule();
1163 if (signal_pending(tsk)) {
1164 ret = -EINTR;
1165 break;
1166 }
1167 /*ret = aio_read_evt(ctx, &ent);*/
1168 } while (1) ;
1169
1170 set_task_state(tsk, TASK_RUNNING);
1171 remove_wait_queue(&ctx->wait, &wait);
1172
1173 if (unlikely(ret <= 0))
1174 break;
1175
1176 ret = -EFAULT;
1177 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1178 dprintk("aio: lost an event due to EFAULT.\n");
1179 break;
1180 }
1181
1182 /* Good, event copied to userland, update counts. */
1183 event ++;
1184 i ++;
1185 }
1186
1187 if (timeout)
1188 clear_timeout(&to);
1189out:
1190 destroy_timer_on_stack(&to.timer);
1191 return i ? i : ret;
1192}
1193
1194/* Take an ioctx and remove it from the list of ioctx's. Protects
1195 * against races with itself via ->dead.
1196 */
1197static void io_destroy(struct kioctx *ioctx)
1198{
1199 struct mm_struct *mm = current->mm;
1200 int was_dead;
1201
1202 /* delete the entry from the list is someone else hasn't already */
1203 spin_lock(&mm->ioctx_lock);
1204 was_dead = ioctx->dead;
1205 ioctx->dead = 1;
1206 hlist_del_rcu(&ioctx->list);
1207 spin_unlock(&mm->ioctx_lock);
1208
1209 dprintk("aio_release(%p)\n", ioctx);
1210 if (likely(!was_dead))
1211 put_ioctx(ioctx); /* twice for the list */
1212
1213 aio_cancel_all(ioctx);
1214 wait_for_all_aios(ioctx);
1215
1216 /*
1217 * Wake up any waiters. The setting of ctx->dead must be seen
1218 * by other CPUs at this point. Right now, we rely on the
1219 * locking done by the above calls to ensure this consistency.
1220 */
1221 wake_up_all(&ioctx->wait);
1222 put_ioctx(ioctx); /* once for the lookup */
1223}
1224
1225/* sys_io_setup:
1226 * Create an aio_context capable of receiving at least nr_events.
1227 * ctxp must not point to an aio_context that already exists, and
1228 * must be initialized to 0 prior to the call. On successful
1229 * creation of the aio_context, *ctxp is filled in with the resulting
1230 * handle. May fail with -EINVAL if *ctxp is not initialized,
1231 * if the specified nr_events exceeds internal limits. May fail
1232 * with -EAGAIN if the specified nr_events exceeds the user's limit
1233 * of available events. May fail with -ENOMEM if insufficient kernel
1234 * resources are available. May fail with -EFAULT if an invalid
1235 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1236 * implemented.
1237 */
1238SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1239{
1240 struct kioctx *ioctx = NULL;
1241 unsigned long ctx;
1242 long ret;
1243
1244 ret = get_user(ctx, ctxp);
1245 if (unlikely(ret))
1246 goto out;
1247
1248 ret = -EINVAL;
1249 if (unlikely(ctx || nr_events == 0)) {
1250 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1251 ctx, nr_events);
1252 goto out;
1253 }
1254
1255 ioctx = ioctx_alloc(nr_events);
1256 ret = PTR_ERR(ioctx);
1257 if (!IS_ERR(ioctx)) {
1258 ret = put_user(ioctx->user_id, ctxp);
1259 if (!ret)
1260 return 0;
1261
1262 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1263 io_destroy(ioctx);
1264 }
1265
1266out:
1267 return ret;
1268}
1269
1270/* sys_io_destroy:
1271 * Destroy the aio_context specified. May cancel any outstanding
1272 * AIOs and block on completion. Will fail with -ENOSYS if not
1273 * implemented. May fail with -EINVAL if the context pointed to
1274 * is invalid.
1275 */
1276SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1277{
1278 struct kioctx *ioctx = lookup_ioctx(ctx);
1279 if (likely(NULL != ioctx)) {
1280 io_destroy(ioctx);
1281 return 0;
1282 }
1283 pr_debug("EINVAL: io_destroy: invalid context id\n");
1284 return -EINVAL;
1285}
1286
1287static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1288{
1289 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1290
1291 BUG_ON(ret <= 0);
1292
1293 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1294 ssize_t this = min((ssize_t)iov->iov_len, ret);
1295 iov->iov_base += this;
1296 iov->iov_len -= this;
1297 iocb->ki_left -= this;
1298 ret -= this;
1299 if (iov->iov_len == 0) {
1300 iocb->ki_cur_seg++;
1301 iov++;
1302 }
1303 }
1304
1305 /* the caller should not have done more io than what fit in
1306 * the remaining iovecs */
1307 BUG_ON(ret > 0 && iocb->ki_left == 0);
1308}
1309
1310static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1311{
1312 struct file *file = iocb->ki_filp;
1313 struct address_space *mapping = file->f_mapping;
1314 struct inode *inode = mapping->host;
1315 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1316 unsigned long, loff_t);
1317 ssize_t ret = 0;
1318 unsigned short opcode;
1319
1320 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1321 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1322 rw_op = file->f_op->aio_read;
1323 opcode = IOCB_CMD_PREADV;
1324 } else {
1325 rw_op = file->f_op->aio_write;
1326 opcode = IOCB_CMD_PWRITEV;
1327 }
1328
1329 /* This matches the pread()/pwrite() logic */
1330 if (iocb->ki_pos < 0)
1331 return -EINVAL;
1332
1333 do {
1334 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1335 iocb->ki_nr_segs - iocb->ki_cur_seg,
1336 iocb->ki_pos);
1337 if (ret > 0)
1338 aio_advance_iovec(iocb, ret);
1339
1340 /* retry all partial writes. retry partial reads as long as its a
1341 * regular file. */
1342 } while (ret > 0 && iocb->ki_left > 0 &&
1343 (opcode == IOCB_CMD_PWRITEV ||
1344 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1345
1346 /* This means we must have transferred all that we could */
1347 /* No need to retry anymore */
1348 if ((ret == 0) || (iocb->ki_left == 0))
1349 ret = iocb->ki_nbytes - iocb->ki_left;
1350
1351 /* If we managed to write some out we return that, rather than
1352 * the eventual error. */
1353 if (opcode == IOCB_CMD_PWRITEV
1354 && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
1355 && iocb->ki_nbytes - iocb->ki_left)
1356 ret = iocb->ki_nbytes - iocb->ki_left;
1357
1358 return ret;
1359}
1360
1361static ssize_t aio_fdsync(struct kiocb *iocb)
1362{
1363 struct file *file = iocb->ki_filp;
1364 ssize_t ret = -EINVAL;
1365
1366 if (file->f_op->aio_fsync)
1367 ret = file->f_op->aio_fsync(iocb, 1);
1368 return ret;
1369}
1370
1371static ssize_t aio_fsync(struct kiocb *iocb)
1372{
1373 struct file *file = iocb->ki_filp;
1374 ssize_t ret = -EINVAL;
1375
1376 if (file->f_op->aio_fsync)
1377 ret = file->f_op->aio_fsync(iocb, 0);
1378 return ret;
1379}
1380
1381static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
1382{
1383 ssize_t ret;
1384
1385#ifdef CONFIG_COMPAT
1386 if (compat)
1387 ret = compat_rw_copy_check_uvector(type,
1388 (struct compat_iovec __user *)kiocb->ki_buf,
1389 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1390 &kiocb->ki_iovec);
1391 else
1392#endif
1393 ret = rw_copy_check_uvector(type,
1394 (struct iovec __user *)kiocb->ki_buf,
1395 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1396 &kiocb->ki_iovec);
1397 if (ret < 0)
1398 goto out;
1399
1400 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1401 kiocb->ki_cur_seg = 0;
1402 /* ki_nbytes/left now reflect bytes instead of segs */
1403 kiocb->ki_nbytes = ret;
1404 kiocb->ki_left = ret;
1405
1406 ret = 0;
1407out:
1408 return ret;
1409}
1410
1411static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
1412{
1413 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1414 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1415 kiocb->ki_iovec->iov_len = kiocb->ki_left;
1416 kiocb->ki_nr_segs = 1;
1417 kiocb->ki_cur_seg = 0;
1418 return 0;
1419}
1420
1421/*
1422 * aio_setup_iocb:
1423 * Performs the initial checks and aio retry method
1424 * setup for the kiocb at the time of io submission.
1425 */
1426static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
1427{
1428 struct file *file = kiocb->ki_filp;
1429 ssize_t ret = 0;
1430
1431 switch (kiocb->ki_opcode) {
1432 case IOCB_CMD_PREAD:
1433 ret = -EBADF;
1434 if (unlikely(!(file->f_mode & FMODE_READ)))
1435 break;
1436 ret = -EFAULT;
1437 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1438 kiocb->ki_left)))
1439 break;
1440 ret = security_file_permission(file, MAY_READ);
1441 if (unlikely(ret))
1442 break;
1443 ret = aio_setup_single_vector(kiocb);
1444 if (ret)
1445 break;
1446 ret = -EINVAL;
1447 if (file->f_op->aio_read)
1448 kiocb->ki_retry = aio_rw_vect_retry;
1449 break;
1450 case IOCB_CMD_PWRITE:
1451 ret = -EBADF;
1452 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1453 break;
1454 ret = -EFAULT;
1455 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1456 kiocb->ki_left)))
1457 break;
1458 ret = security_file_permission(file, MAY_WRITE);
1459 if (unlikely(ret))
1460 break;
1461 ret = aio_setup_single_vector(kiocb);
1462 if (ret)
1463 break;
1464 ret = -EINVAL;
1465 if (file->f_op->aio_write)
1466 kiocb->ki_retry = aio_rw_vect_retry;
1467 break;
1468 case IOCB_CMD_PREADV:
1469 ret = -EBADF;
1470 if (unlikely(!(file->f_mode & FMODE_READ)))
1471 break;
1472 ret = security_file_permission(file, MAY_READ);
1473 if (unlikely(ret))
1474 break;
1475 ret = aio_setup_vectored_rw(READ, kiocb, compat);
1476 if (ret)
1477 break;
1478 ret = -EINVAL;
1479 if (file->f_op->aio_read)
1480 kiocb->ki_retry = aio_rw_vect_retry;
1481 break;
1482 case IOCB_CMD_PWRITEV:
1483 ret = -EBADF;
1484 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1485 break;
1486 ret = security_file_permission(file, MAY_WRITE);
1487 if (unlikely(ret))
1488 break;
1489 ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1490 if (ret)
1491 break;
1492 ret = -EINVAL;
1493 if (file->f_op->aio_write)
1494 kiocb->ki_retry = aio_rw_vect_retry;
1495 break;
1496 case IOCB_CMD_FDSYNC:
1497 ret = -EINVAL;
1498 if (file->f_op->aio_fsync)
1499 kiocb->ki_retry = aio_fdsync;
1500 break;
1501 case IOCB_CMD_FSYNC:
1502 ret = -EINVAL;
1503 if (file->f_op->aio_fsync)
1504 kiocb->ki_retry = aio_fsync;
1505 break;
1506 default:
1507 dprintk("EINVAL: io_submit: no operation provided\n");
1508 ret = -EINVAL;
1509 }
1510
1511 if (!kiocb->ki_retry)
1512 return ret;
1513
1514 return 0;
1515}
1516
1517static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1518 struct iocb *iocb, bool compat)
1519{
1520 struct kiocb *req;
1521 struct file *file;
1522 ssize_t ret;
1523
1524 /* enforce forwards compatibility on users */
1525 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1526 pr_debug("EINVAL: io_submit: reserve field set\n");
1527 return -EINVAL;
1528 }
1529
1530 /* prevent overflows */
1531 if (unlikely(
1532 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1533 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1534 ((ssize_t)iocb->aio_nbytes < 0)
1535 )) {
1536 pr_debug("EINVAL: io_submit: overflow check\n");
1537 return -EINVAL;
1538 }
1539
1540 file = fget(iocb->aio_fildes);
1541 if (unlikely(!file))
1542 return -EBADF;
1543
1544 req = aio_get_req(ctx); /* returns with 2 references to req */
1545 if (unlikely(!req)) {
1546 fput(file);
1547 return -EAGAIN;
1548 }
1549 req->ki_filp = file;
1550 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1551 /*
1552 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1553 * instance of the file* now. The file descriptor must be
1554 * an eventfd() fd, and will be signaled for each completed
1555 * event using the eventfd_signal() function.
1556 */
1557 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1558 if (IS_ERR(req->ki_eventfd)) {
1559 ret = PTR_ERR(req->ki_eventfd);
1560 req->ki_eventfd = NULL;
1561 goto out_put_req;
1562 }
1563 }
1564
1565 ret = put_user(req->ki_key, &user_iocb->aio_key);
1566 if (unlikely(ret)) {
1567 dprintk("EFAULT: aio_key\n");
1568 goto out_put_req;
1569 }
1570
1571 req->ki_obj.user = user_iocb;
1572 req->ki_user_data = iocb->aio_data;
1573 req->ki_pos = iocb->aio_offset;
1574
1575 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1576 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1577 req->ki_opcode = iocb->aio_lio_opcode;
1578
1579 ret = aio_setup_iocb(req, compat);
1580
1581 if (ret)
1582 goto out_put_req;
1583
1584 spin_lock_irq(&ctx->ctx_lock);
1585 /*
1586 * We could have raced with io_destroy() and are currently holding a
1587 * reference to ctx which should be destroyed. We cannot submit IO
1588 * since ctx gets freed as soon as io_submit() puts its reference. The
1589 * check here is reliable: io_destroy() sets ctx->dead before waiting
1590 * for outstanding IO and the barrier between these two is realized by
1591 * unlock of mm->ioctx_lock and lock of ctx->ctx_lock. Analogously we
1592 * increment ctx->reqs_active before checking for ctx->dead and the
1593 * barrier is realized by unlock and lock of ctx->ctx_lock. Thus if we
1594 * don't see ctx->dead set here, io_destroy() waits for our IO to
1595 * finish.
1596 */
1597 if (ctx->dead) {
1598 spin_unlock_irq(&ctx->ctx_lock);
1599 ret = -EINVAL;
1600 goto out_put_req;
1601 }
1602 aio_run_iocb(req);
1603 if (!list_empty(&ctx->run_list)) {
1604 /* drain the run list */
1605 while (__aio_run_iocbs(ctx))
1606 ;
1607 }
1608 spin_unlock_irq(&ctx->ctx_lock);
1609
1610 aio_put_req(req); /* drop extra ref to req */
1611 return 0;
1612
1613out_put_req:
1614 aio_put_req(req); /* drop extra ref to req */
1615 aio_put_req(req); /* drop i/o ref to req */
1616 return ret;
1617}
1618
1619long do_io_submit(aio_context_t ctx_id, long nr,
1620 struct iocb __user *__user *iocbpp, bool compat)
1621{
1622 struct kioctx *ctx;
1623 long ret = 0;
1624 int i;
1625 struct blk_plug plug;
1626
1627 if (unlikely(nr < 0))
1628 return -EINVAL;
1629
1630 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1631 nr = LONG_MAX/sizeof(*iocbpp);
1632
1633 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1634 return -EFAULT;
1635
1636 ctx = lookup_ioctx(ctx_id);
1637 if (unlikely(!ctx)) {
1638 pr_debug("EINVAL: io_submit: invalid context id\n");
1639 return -EINVAL;
1640 }
1641
1642 blk_start_plug(&plug);
1643
1644 /*
1645 * AKPM: should this return a partial result if some of the IOs were
1646 * successfully submitted?
1647 */
1648 for (i=0; i<nr; i++) {
1649 struct iocb __user *user_iocb;
1650 struct iocb tmp;
1651
1652 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1653 ret = -EFAULT;
1654 break;
1655 }
1656
1657 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1658 ret = -EFAULT;
1659 break;
1660 }
1661
1662 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1663 if (ret)
1664 break;
1665 }
1666 blk_finish_plug(&plug);
1667
1668 put_ioctx(ctx);
1669 return i ? i : ret;
1670}
1671
1672/* sys_io_submit:
1673 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1674 * the number of iocbs queued. May return -EINVAL if the aio_context
1675 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1676 * *iocbpp[0] is not properly initialized, if the operation specified
1677 * is invalid for the file descriptor in the iocb. May fail with
1678 * -EFAULT if any of the data structures point to invalid data. May
1679 * fail with -EBADF if the file descriptor specified in the first
1680 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1681 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1682 * fail with -ENOSYS if not implemented.
1683 */
1684SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1685 struct iocb __user * __user *, iocbpp)
1686{
1687 return do_io_submit(ctx_id, nr, iocbpp, 0);
1688}
1689
1690/* lookup_kiocb
1691 * Finds a given iocb for cancellation.
1692 */
1693static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1694 u32 key)
1695{
1696 struct list_head *pos;
1697
1698 assert_spin_locked(&ctx->ctx_lock);
1699
1700 /* TODO: use a hash or array, this sucks. */
1701 list_for_each(pos, &ctx->active_reqs) {
1702 struct kiocb *kiocb = list_kiocb(pos);
1703 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1704 return kiocb;
1705 }
1706 return NULL;
1707}
1708
1709/* sys_io_cancel:
1710 * Attempts to cancel an iocb previously passed to io_submit. If
1711 * the operation is successfully cancelled, the resulting event is
1712 * copied into the memory pointed to by result without being placed
1713 * into the completion queue and 0 is returned. May fail with
1714 * -EFAULT if any of the data structures pointed to are invalid.
1715 * May fail with -EINVAL if aio_context specified by ctx_id is
1716 * invalid. May fail with -EAGAIN if the iocb specified was not
1717 * cancelled. Will fail with -ENOSYS if not implemented.
1718 */
1719SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1720 struct io_event __user *, result)
1721{
1722 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1723 struct kioctx *ctx;
1724 struct kiocb *kiocb;
1725 u32 key;
1726 int ret;
1727
1728 ret = get_user(key, &iocb->aio_key);
1729 if (unlikely(ret))
1730 return -EFAULT;
1731
1732 ctx = lookup_ioctx(ctx_id);
1733 if (unlikely(!ctx))
1734 return -EINVAL;
1735
1736 spin_lock_irq(&ctx->ctx_lock);
1737 ret = -EAGAIN;
1738 kiocb = lookup_kiocb(ctx, iocb, key);
1739 if (kiocb && kiocb->ki_cancel) {
1740 cancel = kiocb->ki_cancel;
1741 kiocb->ki_users ++;
1742 kiocbSetCancelled(kiocb);
1743 } else
1744 cancel = NULL;
1745 spin_unlock_irq(&ctx->ctx_lock);
1746
1747 if (NULL != cancel) {
1748 struct io_event tmp;
1749 pr_debug("calling cancel\n");
1750 memset(&tmp, 0, sizeof(tmp));
1751 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1752 tmp.data = kiocb->ki_user_data;
1753 ret = cancel(kiocb, &tmp);
1754 if (!ret) {
1755 /* Cancellation succeeded -- copy the result
1756 * into the user's buffer.
1757 */
1758 if (copy_to_user(result, &tmp, sizeof(tmp)))
1759 ret = -EFAULT;
1760 }
1761 } else
1762 ret = -EINVAL;
1763
1764 put_ioctx(ctx);
1765
1766 return ret;
1767}
1768
1769/* io_getevents:
1770 * Attempts to read at least min_nr events and up to nr events from
1771 * the completion queue for the aio_context specified by ctx_id. If
1772 * it succeeds, the number of read events is returned. May fail with
1773 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1774 * out of range, if timeout is out of range. May fail with -EFAULT
1775 * if any of the memory specified is invalid. May return 0 or
1776 * < min_nr if the timeout specified by timeout has elapsed
1777 * before sufficient events are available, where timeout == NULL
1778 * specifies an infinite timeout. Note that the timeout pointed to by
1779 * timeout is relative and will be updated if not NULL and the
1780 * operation blocks. Will fail with -ENOSYS if not implemented.
1781 */
1782SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1783 long, min_nr,
1784 long, nr,
1785 struct io_event __user *, events,
1786 struct timespec __user *, timeout)
1787{
1788 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1789 long ret = -EINVAL;
1790
1791 if (likely(ioctx)) {
1792 if (likely(min_nr <= nr && min_nr >= 0))
1793 ret = read_events(ioctx, min_nr, nr, events, timeout);
1794 put_ioctx(ioctx);
1795 }
1796
1797 asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);
1798 return ret;
1799}
1/*
2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <bcrl@kvack.org>
4 *
5 * Implements an efficient asynchronous io interface.
6 *
7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved.
8 * Copyright 2018 Christoph Hellwig.
9 *
10 * See ../COPYING for licensing terms.
11 */
12#define pr_fmt(fmt) "%s: " fmt, __func__
13
14#include <linux/kernel.h>
15#include <linux/init.h>
16#include <linux/errno.h>
17#include <linux/time.h>
18#include <linux/aio_abi.h>
19#include <linux/export.h>
20#include <linux/syscalls.h>
21#include <linux/backing-dev.h>
22#include <linux/refcount.h>
23#include <linux/uio.h>
24
25#include <linux/sched/signal.h>
26#include <linux/fs.h>
27#include <linux/file.h>
28#include <linux/mm.h>
29#include <linux/mman.h>
30#include <linux/mmu_context.h>
31#include <linux/percpu.h>
32#include <linux/slab.h>
33#include <linux/timer.h>
34#include <linux/aio.h>
35#include <linux/highmem.h>
36#include <linux/workqueue.h>
37#include <linux/security.h>
38#include <linux/eventfd.h>
39#include <linux/blkdev.h>
40#include <linux/compat.h>
41#include <linux/migrate.h>
42#include <linux/ramfs.h>
43#include <linux/percpu-refcount.h>
44#include <linux/mount.h>
45#include <linux/pseudo_fs.h>
46
47#include <asm/kmap_types.h>
48#include <linux/uaccess.h>
49#include <linux/nospec.h>
50
51#include "internal.h"
52
53#define KIOCB_KEY 0
54
55#define AIO_RING_MAGIC 0xa10a10a1
56#define AIO_RING_COMPAT_FEATURES 1
57#define AIO_RING_INCOMPAT_FEATURES 0
58struct aio_ring {
59 unsigned id; /* kernel internal index number */
60 unsigned nr; /* number of io_events */
61 unsigned head; /* Written to by userland or under ring_lock
62 * mutex by aio_read_events_ring(). */
63 unsigned tail;
64
65 unsigned magic;
66 unsigned compat_features;
67 unsigned incompat_features;
68 unsigned header_length; /* size of aio_ring */
69
70
71 struct io_event io_events[0];
72}; /* 128 bytes + ring size */
73
74/*
75 * Plugging is meant to work with larger batches of IOs. If we don't
76 * have more than the below, then don't bother setting up a plug.
77 */
78#define AIO_PLUG_THRESHOLD 2
79
80#define AIO_RING_PAGES 8
81
82struct kioctx_table {
83 struct rcu_head rcu;
84 unsigned nr;
85 struct kioctx __rcu *table[];
86};
87
88struct kioctx_cpu {
89 unsigned reqs_available;
90};
91
92struct ctx_rq_wait {
93 struct completion comp;
94 atomic_t count;
95};
96
97struct kioctx {
98 struct percpu_ref users;
99 atomic_t dead;
100
101 struct percpu_ref reqs;
102
103 unsigned long user_id;
104
105 struct __percpu kioctx_cpu *cpu;
106
107 /*
108 * For percpu reqs_available, number of slots we move to/from global
109 * counter at a time:
110 */
111 unsigned req_batch;
112 /*
113 * This is what userspace passed to io_setup(), it's not used for
114 * anything but counting against the global max_reqs quota.
115 *
116 * The real limit is nr_events - 1, which will be larger (see
117 * aio_setup_ring())
118 */
119 unsigned max_reqs;
120
121 /* Size of ringbuffer, in units of struct io_event */
122 unsigned nr_events;
123
124 unsigned long mmap_base;
125 unsigned long mmap_size;
126
127 struct page **ring_pages;
128 long nr_pages;
129
130 struct rcu_work free_rwork; /* see free_ioctx() */
131
132 /*
133 * signals when all in-flight requests are done
134 */
135 struct ctx_rq_wait *rq_wait;
136
137 struct {
138 /*
139 * This counts the number of available slots in the ringbuffer,
140 * so we avoid overflowing it: it's decremented (if positive)
141 * when allocating a kiocb and incremented when the resulting
142 * io_event is pulled off the ringbuffer.
143 *
144 * We batch accesses to it with a percpu version.
145 */
146 atomic_t reqs_available;
147 } ____cacheline_aligned_in_smp;
148
149 struct {
150 spinlock_t ctx_lock;
151 struct list_head active_reqs; /* used for cancellation */
152 } ____cacheline_aligned_in_smp;
153
154 struct {
155 struct mutex ring_lock;
156 wait_queue_head_t wait;
157 } ____cacheline_aligned_in_smp;
158
159 struct {
160 unsigned tail;
161 unsigned completed_events;
162 spinlock_t completion_lock;
163 } ____cacheline_aligned_in_smp;
164
165 struct page *internal_pages[AIO_RING_PAGES];
166 struct file *aio_ring_file;
167
168 unsigned id;
169};
170
171/*
172 * First field must be the file pointer in all the
173 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
174 */
175struct fsync_iocb {
176 struct file *file;
177 struct work_struct work;
178 bool datasync;
179};
180
181struct poll_iocb {
182 struct file *file;
183 struct wait_queue_head *head;
184 __poll_t events;
185 bool done;
186 bool cancelled;
187 struct wait_queue_entry wait;
188 struct work_struct work;
189};
190
191/*
192 * NOTE! Each of the iocb union members has the file pointer
193 * as the first entry in their struct definition. So you can
194 * access the file pointer through any of the sub-structs,
195 * or directly as just 'ki_filp' in this struct.
196 */
197struct aio_kiocb {
198 union {
199 struct file *ki_filp;
200 struct kiocb rw;
201 struct fsync_iocb fsync;
202 struct poll_iocb poll;
203 };
204
205 struct kioctx *ki_ctx;
206 kiocb_cancel_fn *ki_cancel;
207
208 struct io_event ki_res;
209
210 struct list_head ki_list; /* the aio core uses this
211 * for cancellation */
212 refcount_t ki_refcnt;
213
214 /*
215 * If the aio_resfd field of the userspace iocb is not zero,
216 * this is the underlying eventfd context to deliver events to.
217 */
218 struct eventfd_ctx *ki_eventfd;
219};
220
221/*------ sysctl variables----*/
222static DEFINE_SPINLOCK(aio_nr_lock);
223unsigned long aio_nr; /* current system wide number of aio requests */
224unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
225/*----end sysctl variables---*/
226
227static struct kmem_cache *kiocb_cachep;
228static struct kmem_cache *kioctx_cachep;
229
230static struct vfsmount *aio_mnt;
231
232static const struct file_operations aio_ring_fops;
233static const struct address_space_operations aio_ctx_aops;
234
235static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
236{
237 struct file *file;
238 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
239 if (IS_ERR(inode))
240 return ERR_CAST(inode);
241
242 inode->i_mapping->a_ops = &aio_ctx_aops;
243 inode->i_mapping->private_data = ctx;
244 inode->i_size = PAGE_SIZE * nr_pages;
245
246 file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
247 O_RDWR, &aio_ring_fops);
248 if (IS_ERR(file))
249 iput(inode);
250 return file;
251}
252
253static int aio_init_fs_context(struct fs_context *fc)
254{
255 if (!init_pseudo(fc, AIO_RING_MAGIC))
256 return -ENOMEM;
257 fc->s_iflags |= SB_I_NOEXEC;
258 return 0;
259}
260
261/* aio_setup
262 * Creates the slab caches used by the aio routines, panic on
263 * failure as this is done early during the boot sequence.
264 */
265static int __init aio_setup(void)
266{
267 static struct file_system_type aio_fs = {
268 .name = "aio",
269 .init_fs_context = aio_init_fs_context,
270 .kill_sb = kill_anon_super,
271 };
272 aio_mnt = kern_mount(&aio_fs);
273 if (IS_ERR(aio_mnt))
274 panic("Failed to create aio fs mount.");
275
276 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
277 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
278 return 0;
279}
280__initcall(aio_setup);
281
282static void put_aio_ring_file(struct kioctx *ctx)
283{
284 struct file *aio_ring_file = ctx->aio_ring_file;
285 struct address_space *i_mapping;
286
287 if (aio_ring_file) {
288 truncate_setsize(file_inode(aio_ring_file), 0);
289
290 /* Prevent further access to the kioctx from migratepages */
291 i_mapping = aio_ring_file->f_mapping;
292 spin_lock(&i_mapping->private_lock);
293 i_mapping->private_data = NULL;
294 ctx->aio_ring_file = NULL;
295 spin_unlock(&i_mapping->private_lock);
296
297 fput(aio_ring_file);
298 }
299}
300
301static void aio_free_ring(struct kioctx *ctx)
302{
303 int i;
304
305 /* Disconnect the kiotx from the ring file. This prevents future
306 * accesses to the kioctx from page migration.
307 */
308 put_aio_ring_file(ctx);
309
310 for (i = 0; i < ctx->nr_pages; i++) {
311 struct page *page;
312 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
313 page_count(ctx->ring_pages[i]));
314 page = ctx->ring_pages[i];
315 if (!page)
316 continue;
317 ctx->ring_pages[i] = NULL;
318 put_page(page);
319 }
320
321 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
322 kfree(ctx->ring_pages);
323 ctx->ring_pages = NULL;
324 }
325}
326
327static int aio_ring_mremap(struct vm_area_struct *vma)
328{
329 struct file *file = vma->vm_file;
330 struct mm_struct *mm = vma->vm_mm;
331 struct kioctx_table *table;
332 int i, res = -EINVAL;
333
334 spin_lock(&mm->ioctx_lock);
335 rcu_read_lock();
336 table = rcu_dereference(mm->ioctx_table);
337 for (i = 0; i < table->nr; i++) {
338 struct kioctx *ctx;
339
340 ctx = rcu_dereference(table->table[i]);
341 if (ctx && ctx->aio_ring_file == file) {
342 if (!atomic_read(&ctx->dead)) {
343 ctx->user_id = ctx->mmap_base = vma->vm_start;
344 res = 0;
345 }
346 break;
347 }
348 }
349
350 rcu_read_unlock();
351 spin_unlock(&mm->ioctx_lock);
352 return res;
353}
354
355static const struct vm_operations_struct aio_ring_vm_ops = {
356 .mremap = aio_ring_mremap,
357#if IS_ENABLED(CONFIG_MMU)
358 .fault = filemap_fault,
359 .map_pages = filemap_map_pages,
360 .page_mkwrite = filemap_page_mkwrite,
361#endif
362};
363
364static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
365{
366 vma->vm_flags |= VM_DONTEXPAND;
367 vma->vm_ops = &aio_ring_vm_ops;
368 return 0;
369}
370
371static const struct file_operations aio_ring_fops = {
372 .mmap = aio_ring_mmap,
373};
374
375#if IS_ENABLED(CONFIG_MIGRATION)
376static int aio_migratepage(struct address_space *mapping, struct page *new,
377 struct page *old, enum migrate_mode mode)
378{
379 struct kioctx *ctx;
380 unsigned long flags;
381 pgoff_t idx;
382 int rc;
383
384 /*
385 * We cannot support the _NO_COPY case here, because copy needs to
386 * happen under the ctx->completion_lock. That does not work with the
387 * migration workflow of MIGRATE_SYNC_NO_COPY.
388 */
389 if (mode == MIGRATE_SYNC_NO_COPY)
390 return -EINVAL;
391
392 rc = 0;
393
394 /* mapping->private_lock here protects against the kioctx teardown. */
395 spin_lock(&mapping->private_lock);
396 ctx = mapping->private_data;
397 if (!ctx) {
398 rc = -EINVAL;
399 goto out;
400 }
401
402 /* The ring_lock mutex. The prevents aio_read_events() from writing
403 * to the ring's head, and prevents page migration from mucking in
404 * a partially initialized kiotx.
405 */
406 if (!mutex_trylock(&ctx->ring_lock)) {
407 rc = -EAGAIN;
408 goto out;
409 }
410
411 idx = old->index;
412 if (idx < (pgoff_t)ctx->nr_pages) {
413 /* Make sure the old page hasn't already been changed */
414 if (ctx->ring_pages[idx] != old)
415 rc = -EAGAIN;
416 } else
417 rc = -EINVAL;
418
419 if (rc != 0)
420 goto out_unlock;
421
422 /* Writeback must be complete */
423 BUG_ON(PageWriteback(old));
424 get_page(new);
425
426 rc = migrate_page_move_mapping(mapping, new, old, 1);
427 if (rc != MIGRATEPAGE_SUCCESS) {
428 put_page(new);
429 goto out_unlock;
430 }
431
432 /* Take completion_lock to prevent other writes to the ring buffer
433 * while the old page is copied to the new. This prevents new
434 * events from being lost.
435 */
436 spin_lock_irqsave(&ctx->completion_lock, flags);
437 migrate_page_copy(new, old);
438 BUG_ON(ctx->ring_pages[idx] != old);
439 ctx->ring_pages[idx] = new;
440 spin_unlock_irqrestore(&ctx->completion_lock, flags);
441
442 /* The old page is no longer accessible. */
443 put_page(old);
444
445out_unlock:
446 mutex_unlock(&ctx->ring_lock);
447out:
448 spin_unlock(&mapping->private_lock);
449 return rc;
450}
451#endif
452
453static const struct address_space_operations aio_ctx_aops = {
454 .set_page_dirty = __set_page_dirty_no_writeback,
455#if IS_ENABLED(CONFIG_MIGRATION)
456 .migratepage = aio_migratepage,
457#endif
458};
459
460static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
461{
462 struct aio_ring *ring;
463 struct mm_struct *mm = current->mm;
464 unsigned long size, unused;
465 int nr_pages;
466 int i;
467 struct file *file;
468
469 /* Compensate for the ring buffer's head/tail overlap entry */
470 nr_events += 2; /* 1 is required, 2 for good luck */
471
472 size = sizeof(struct aio_ring);
473 size += sizeof(struct io_event) * nr_events;
474
475 nr_pages = PFN_UP(size);
476 if (nr_pages < 0)
477 return -EINVAL;
478
479 file = aio_private_file(ctx, nr_pages);
480 if (IS_ERR(file)) {
481 ctx->aio_ring_file = NULL;
482 return -ENOMEM;
483 }
484
485 ctx->aio_ring_file = file;
486 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
487 / sizeof(struct io_event);
488
489 ctx->ring_pages = ctx->internal_pages;
490 if (nr_pages > AIO_RING_PAGES) {
491 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
492 GFP_KERNEL);
493 if (!ctx->ring_pages) {
494 put_aio_ring_file(ctx);
495 return -ENOMEM;
496 }
497 }
498
499 for (i = 0; i < nr_pages; i++) {
500 struct page *page;
501 page = find_or_create_page(file->f_mapping,
502 i, GFP_HIGHUSER | __GFP_ZERO);
503 if (!page)
504 break;
505 pr_debug("pid(%d) page[%d]->count=%d\n",
506 current->pid, i, page_count(page));
507 SetPageUptodate(page);
508 unlock_page(page);
509
510 ctx->ring_pages[i] = page;
511 }
512 ctx->nr_pages = i;
513
514 if (unlikely(i != nr_pages)) {
515 aio_free_ring(ctx);
516 return -ENOMEM;
517 }
518
519 ctx->mmap_size = nr_pages * PAGE_SIZE;
520 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
521
522 if (down_write_killable(&mm->mmap_sem)) {
523 ctx->mmap_size = 0;
524 aio_free_ring(ctx);
525 return -EINTR;
526 }
527
528 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
529 PROT_READ | PROT_WRITE,
530 MAP_SHARED, 0, &unused, NULL);
531 up_write(&mm->mmap_sem);
532 if (IS_ERR((void *)ctx->mmap_base)) {
533 ctx->mmap_size = 0;
534 aio_free_ring(ctx);
535 return -ENOMEM;
536 }
537
538 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
539
540 ctx->user_id = ctx->mmap_base;
541 ctx->nr_events = nr_events; /* trusted copy */
542
543 ring = kmap_atomic(ctx->ring_pages[0]);
544 ring->nr = nr_events; /* user copy */
545 ring->id = ~0U;
546 ring->head = ring->tail = 0;
547 ring->magic = AIO_RING_MAGIC;
548 ring->compat_features = AIO_RING_COMPAT_FEATURES;
549 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
550 ring->header_length = sizeof(struct aio_ring);
551 kunmap_atomic(ring);
552 flush_dcache_page(ctx->ring_pages[0]);
553
554 return 0;
555}
556
557#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
558#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
559#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
560
561void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
562{
563 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
564 struct kioctx *ctx = req->ki_ctx;
565 unsigned long flags;
566
567 if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
568 return;
569
570 spin_lock_irqsave(&ctx->ctx_lock, flags);
571 list_add_tail(&req->ki_list, &ctx->active_reqs);
572 req->ki_cancel = cancel;
573 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
574}
575EXPORT_SYMBOL(kiocb_set_cancel_fn);
576
577/*
578 * free_ioctx() should be RCU delayed to synchronize against the RCU
579 * protected lookup_ioctx() and also needs process context to call
580 * aio_free_ring(). Use rcu_work.
581 */
582static void free_ioctx(struct work_struct *work)
583{
584 struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
585 free_rwork);
586 pr_debug("freeing %p\n", ctx);
587
588 aio_free_ring(ctx);
589 free_percpu(ctx->cpu);
590 percpu_ref_exit(&ctx->reqs);
591 percpu_ref_exit(&ctx->users);
592 kmem_cache_free(kioctx_cachep, ctx);
593}
594
595static void free_ioctx_reqs(struct percpu_ref *ref)
596{
597 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
598
599 /* At this point we know that there are no any in-flight requests */
600 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
601 complete(&ctx->rq_wait->comp);
602
603 /* Synchronize against RCU protected table->table[] dereferences */
604 INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
605 queue_rcu_work(system_wq, &ctx->free_rwork);
606}
607
608/*
609 * When this function runs, the kioctx has been removed from the "hash table"
610 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
611 * now it's safe to cancel any that need to be.
612 */
613static void free_ioctx_users(struct percpu_ref *ref)
614{
615 struct kioctx *ctx = container_of(ref, struct kioctx, users);
616 struct aio_kiocb *req;
617
618 spin_lock_irq(&ctx->ctx_lock);
619
620 while (!list_empty(&ctx->active_reqs)) {
621 req = list_first_entry(&ctx->active_reqs,
622 struct aio_kiocb, ki_list);
623 req->ki_cancel(&req->rw);
624 list_del_init(&req->ki_list);
625 }
626
627 spin_unlock_irq(&ctx->ctx_lock);
628
629 percpu_ref_kill(&ctx->reqs);
630 percpu_ref_put(&ctx->reqs);
631}
632
633static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
634{
635 unsigned i, new_nr;
636 struct kioctx_table *table, *old;
637 struct aio_ring *ring;
638
639 spin_lock(&mm->ioctx_lock);
640 table = rcu_dereference_raw(mm->ioctx_table);
641
642 while (1) {
643 if (table)
644 for (i = 0; i < table->nr; i++)
645 if (!rcu_access_pointer(table->table[i])) {
646 ctx->id = i;
647 rcu_assign_pointer(table->table[i], ctx);
648 spin_unlock(&mm->ioctx_lock);
649
650 /* While kioctx setup is in progress,
651 * we are protected from page migration
652 * changes ring_pages by ->ring_lock.
653 */
654 ring = kmap_atomic(ctx->ring_pages[0]);
655 ring->id = ctx->id;
656 kunmap_atomic(ring);
657 return 0;
658 }
659
660 new_nr = (table ? table->nr : 1) * 4;
661 spin_unlock(&mm->ioctx_lock);
662
663 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
664 new_nr, GFP_KERNEL);
665 if (!table)
666 return -ENOMEM;
667
668 table->nr = new_nr;
669
670 spin_lock(&mm->ioctx_lock);
671 old = rcu_dereference_raw(mm->ioctx_table);
672
673 if (!old) {
674 rcu_assign_pointer(mm->ioctx_table, table);
675 } else if (table->nr > old->nr) {
676 memcpy(table->table, old->table,
677 old->nr * sizeof(struct kioctx *));
678
679 rcu_assign_pointer(mm->ioctx_table, table);
680 kfree_rcu(old, rcu);
681 } else {
682 kfree(table);
683 table = old;
684 }
685 }
686}
687
688static void aio_nr_sub(unsigned nr)
689{
690 spin_lock(&aio_nr_lock);
691 if (WARN_ON(aio_nr - nr > aio_nr))
692 aio_nr = 0;
693 else
694 aio_nr -= nr;
695 spin_unlock(&aio_nr_lock);
696}
697
698/* ioctx_alloc
699 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
700 */
701static struct kioctx *ioctx_alloc(unsigned nr_events)
702{
703 struct mm_struct *mm = current->mm;
704 struct kioctx *ctx;
705 int err = -ENOMEM;
706
707 /*
708 * Store the original nr_events -- what userspace passed to io_setup(),
709 * for counting against the global limit -- before it changes.
710 */
711 unsigned int max_reqs = nr_events;
712
713 /*
714 * We keep track of the number of available ringbuffer slots, to prevent
715 * overflow (reqs_available), and we also use percpu counters for this.
716 *
717 * So since up to half the slots might be on other cpu's percpu counters
718 * and unavailable, double nr_events so userspace sees what they
719 * expected: additionally, we move req_batch slots to/from percpu
720 * counters at a time, so make sure that isn't 0:
721 */
722 nr_events = max(nr_events, num_possible_cpus() * 4);
723 nr_events *= 2;
724
725 /* Prevent overflows */
726 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
727 pr_debug("ENOMEM: nr_events too high\n");
728 return ERR_PTR(-EINVAL);
729 }
730
731 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
732 return ERR_PTR(-EAGAIN);
733
734 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
735 if (!ctx)
736 return ERR_PTR(-ENOMEM);
737
738 ctx->max_reqs = max_reqs;
739
740 spin_lock_init(&ctx->ctx_lock);
741 spin_lock_init(&ctx->completion_lock);
742 mutex_init(&ctx->ring_lock);
743 /* Protect against page migration throughout kiotx setup by keeping
744 * the ring_lock mutex held until setup is complete. */
745 mutex_lock(&ctx->ring_lock);
746 init_waitqueue_head(&ctx->wait);
747
748 INIT_LIST_HEAD(&ctx->active_reqs);
749
750 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
751 goto err;
752
753 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
754 goto err;
755
756 ctx->cpu = alloc_percpu(struct kioctx_cpu);
757 if (!ctx->cpu)
758 goto err;
759
760 err = aio_setup_ring(ctx, nr_events);
761 if (err < 0)
762 goto err;
763
764 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
765 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
766 if (ctx->req_batch < 1)
767 ctx->req_batch = 1;
768
769 /* limit the number of system wide aios */
770 spin_lock(&aio_nr_lock);
771 if (aio_nr + ctx->max_reqs > aio_max_nr ||
772 aio_nr + ctx->max_reqs < aio_nr) {
773 spin_unlock(&aio_nr_lock);
774 err = -EAGAIN;
775 goto err_ctx;
776 }
777 aio_nr += ctx->max_reqs;
778 spin_unlock(&aio_nr_lock);
779
780 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
781 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
782
783 err = ioctx_add_table(ctx, mm);
784 if (err)
785 goto err_cleanup;
786
787 /* Release the ring_lock mutex now that all setup is complete. */
788 mutex_unlock(&ctx->ring_lock);
789
790 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
791 ctx, ctx->user_id, mm, ctx->nr_events);
792 return ctx;
793
794err_cleanup:
795 aio_nr_sub(ctx->max_reqs);
796err_ctx:
797 atomic_set(&ctx->dead, 1);
798 if (ctx->mmap_size)
799 vm_munmap(ctx->mmap_base, ctx->mmap_size);
800 aio_free_ring(ctx);
801err:
802 mutex_unlock(&ctx->ring_lock);
803 free_percpu(ctx->cpu);
804 percpu_ref_exit(&ctx->reqs);
805 percpu_ref_exit(&ctx->users);
806 kmem_cache_free(kioctx_cachep, ctx);
807 pr_debug("error allocating ioctx %d\n", err);
808 return ERR_PTR(err);
809}
810
811/* kill_ioctx
812 * Cancels all outstanding aio requests on an aio context. Used
813 * when the processes owning a context have all exited to encourage
814 * the rapid destruction of the kioctx.
815 */
816static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
817 struct ctx_rq_wait *wait)
818{
819 struct kioctx_table *table;
820
821 spin_lock(&mm->ioctx_lock);
822 if (atomic_xchg(&ctx->dead, 1)) {
823 spin_unlock(&mm->ioctx_lock);
824 return -EINVAL;
825 }
826
827 table = rcu_dereference_raw(mm->ioctx_table);
828 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
829 RCU_INIT_POINTER(table->table[ctx->id], NULL);
830 spin_unlock(&mm->ioctx_lock);
831
832 /* free_ioctx_reqs() will do the necessary RCU synchronization */
833 wake_up_all(&ctx->wait);
834
835 /*
836 * It'd be more correct to do this in free_ioctx(), after all
837 * the outstanding kiocbs have finished - but by then io_destroy
838 * has already returned, so io_setup() could potentially return
839 * -EAGAIN with no ioctxs actually in use (as far as userspace
840 * could tell).
841 */
842 aio_nr_sub(ctx->max_reqs);
843
844 if (ctx->mmap_size)
845 vm_munmap(ctx->mmap_base, ctx->mmap_size);
846
847 ctx->rq_wait = wait;
848 percpu_ref_kill(&ctx->users);
849 return 0;
850}
851
852/*
853 * exit_aio: called when the last user of mm goes away. At this point, there is
854 * no way for any new requests to be submited or any of the io_* syscalls to be
855 * called on the context.
856 *
857 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
858 * them.
859 */
860void exit_aio(struct mm_struct *mm)
861{
862 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
863 struct ctx_rq_wait wait;
864 int i, skipped;
865
866 if (!table)
867 return;
868
869 atomic_set(&wait.count, table->nr);
870 init_completion(&wait.comp);
871
872 skipped = 0;
873 for (i = 0; i < table->nr; ++i) {
874 struct kioctx *ctx =
875 rcu_dereference_protected(table->table[i], true);
876
877 if (!ctx) {
878 skipped++;
879 continue;
880 }
881
882 /*
883 * We don't need to bother with munmap() here - exit_mmap(mm)
884 * is coming and it'll unmap everything. And we simply can't,
885 * this is not necessarily our ->mm.
886 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
887 * that it needs to unmap the area, just set it to 0.
888 */
889 ctx->mmap_size = 0;
890 kill_ioctx(mm, ctx, &wait);
891 }
892
893 if (!atomic_sub_and_test(skipped, &wait.count)) {
894 /* Wait until all IO for the context are done. */
895 wait_for_completion(&wait.comp);
896 }
897
898 RCU_INIT_POINTER(mm->ioctx_table, NULL);
899 kfree(table);
900}
901
902static void put_reqs_available(struct kioctx *ctx, unsigned nr)
903{
904 struct kioctx_cpu *kcpu;
905 unsigned long flags;
906
907 local_irq_save(flags);
908 kcpu = this_cpu_ptr(ctx->cpu);
909 kcpu->reqs_available += nr;
910
911 while (kcpu->reqs_available >= ctx->req_batch * 2) {
912 kcpu->reqs_available -= ctx->req_batch;
913 atomic_add(ctx->req_batch, &ctx->reqs_available);
914 }
915
916 local_irq_restore(flags);
917}
918
919static bool __get_reqs_available(struct kioctx *ctx)
920{
921 struct kioctx_cpu *kcpu;
922 bool ret = false;
923 unsigned long flags;
924
925 local_irq_save(flags);
926 kcpu = this_cpu_ptr(ctx->cpu);
927 if (!kcpu->reqs_available) {
928 int old, avail = atomic_read(&ctx->reqs_available);
929
930 do {
931 if (avail < ctx->req_batch)
932 goto out;
933
934 old = avail;
935 avail = atomic_cmpxchg(&ctx->reqs_available,
936 avail, avail - ctx->req_batch);
937 } while (avail != old);
938
939 kcpu->reqs_available += ctx->req_batch;
940 }
941
942 ret = true;
943 kcpu->reqs_available--;
944out:
945 local_irq_restore(flags);
946 return ret;
947}
948
949/* refill_reqs_available
950 * Updates the reqs_available reference counts used for tracking the
951 * number of free slots in the completion ring. This can be called
952 * from aio_complete() (to optimistically update reqs_available) or
953 * from aio_get_req() (the we're out of events case). It must be
954 * called holding ctx->completion_lock.
955 */
956static void refill_reqs_available(struct kioctx *ctx, unsigned head,
957 unsigned tail)
958{
959 unsigned events_in_ring, completed;
960
961 /* Clamp head since userland can write to it. */
962 head %= ctx->nr_events;
963 if (head <= tail)
964 events_in_ring = tail - head;
965 else
966 events_in_ring = ctx->nr_events - (head - tail);
967
968 completed = ctx->completed_events;
969 if (events_in_ring < completed)
970 completed -= events_in_ring;
971 else
972 completed = 0;
973
974 if (!completed)
975 return;
976
977 ctx->completed_events -= completed;
978 put_reqs_available(ctx, completed);
979}
980
981/* user_refill_reqs_available
982 * Called to refill reqs_available when aio_get_req() encounters an
983 * out of space in the completion ring.
984 */
985static void user_refill_reqs_available(struct kioctx *ctx)
986{
987 spin_lock_irq(&ctx->completion_lock);
988 if (ctx->completed_events) {
989 struct aio_ring *ring;
990 unsigned head;
991
992 /* Access of ring->head may race with aio_read_events_ring()
993 * here, but that's okay since whether we read the old version
994 * or the new version, and either will be valid. The important
995 * part is that head cannot pass tail since we prevent
996 * aio_complete() from updating tail by holding
997 * ctx->completion_lock. Even if head is invalid, the check
998 * against ctx->completed_events below will make sure we do the
999 * safe/right thing.
1000 */
1001 ring = kmap_atomic(ctx->ring_pages[0]);
1002 head = ring->head;
1003 kunmap_atomic(ring);
1004
1005 refill_reqs_available(ctx, head, ctx->tail);
1006 }
1007
1008 spin_unlock_irq(&ctx->completion_lock);
1009}
1010
1011static bool get_reqs_available(struct kioctx *ctx)
1012{
1013 if (__get_reqs_available(ctx))
1014 return true;
1015 user_refill_reqs_available(ctx);
1016 return __get_reqs_available(ctx);
1017}
1018
1019/* aio_get_req
1020 * Allocate a slot for an aio request.
1021 * Returns NULL if no requests are free.
1022 *
1023 * The refcount is initialized to 2 - one for the async op completion,
1024 * one for the synchronous code that does this.
1025 */
1026static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1027{
1028 struct aio_kiocb *req;
1029
1030 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1031 if (unlikely(!req))
1032 return NULL;
1033
1034 if (unlikely(!get_reqs_available(ctx))) {
1035 kmem_cache_free(kiocb_cachep, req);
1036 return NULL;
1037 }
1038
1039 percpu_ref_get(&ctx->reqs);
1040 req->ki_ctx = ctx;
1041 INIT_LIST_HEAD(&req->ki_list);
1042 refcount_set(&req->ki_refcnt, 2);
1043 req->ki_eventfd = NULL;
1044 return req;
1045}
1046
1047static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1048{
1049 struct aio_ring __user *ring = (void __user *)ctx_id;
1050 struct mm_struct *mm = current->mm;
1051 struct kioctx *ctx, *ret = NULL;
1052 struct kioctx_table *table;
1053 unsigned id;
1054
1055 if (get_user(id, &ring->id))
1056 return NULL;
1057
1058 rcu_read_lock();
1059 table = rcu_dereference(mm->ioctx_table);
1060
1061 if (!table || id >= table->nr)
1062 goto out;
1063
1064 id = array_index_nospec(id, table->nr);
1065 ctx = rcu_dereference(table->table[id]);
1066 if (ctx && ctx->user_id == ctx_id) {
1067 if (percpu_ref_tryget_live(&ctx->users))
1068 ret = ctx;
1069 }
1070out:
1071 rcu_read_unlock();
1072 return ret;
1073}
1074
1075static inline void iocb_destroy(struct aio_kiocb *iocb)
1076{
1077 if (iocb->ki_eventfd)
1078 eventfd_ctx_put(iocb->ki_eventfd);
1079 if (iocb->ki_filp)
1080 fput(iocb->ki_filp);
1081 percpu_ref_put(&iocb->ki_ctx->reqs);
1082 kmem_cache_free(kiocb_cachep, iocb);
1083}
1084
1085/* aio_complete
1086 * Called when the io request on the given iocb is complete.
1087 */
1088static void aio_complete(struct aio_kiocb *iocb)
1089{
1090 struct kioctx *ctx = iocb->ki_ctx;
1091 struct aio_ring *ring;
1092 struct io_event *ev_page, *event;
1093 unsigned tail, pos, head;
1094 unsigned long flags;
1095
1096 /*
1097 * Add a completion event to the ring buffer. Must be done holding
1098 * ctx->completion_lock to prevent other code from messing with the tail
1099 * pointer since we might be called from irq context.
1100 */
1101 spin_lock_irqsave(&ctx->completion_lock, flags);
1102
1103 tail = ctx->tail;
1104 pos = tail + AIO_EVENTS_OFFSET;
1105
1106 if (++tail >= ctx->nr_events)
1107 tail = 0;
1108
1109 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1110 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1111
1112 *event = iocb->ki_res;
1113
1114 kunmap_atomic(ev_page);
1115 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1116
1117 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1118 (void __user *)(unsigned long)iocb->ki_res.obj,
1119 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1120
1121 /* after flagging the request as done, we
1122 * must never even look at it again
1123 */
1124 smp_wmb(); /* make event visible before updating tail */
1125
1126 ctx->tail = tail;
1127
1128 ring = kmap_atomic(ctx->ring_pages[0]);
1129 head = ring->head;
1130 ring->tail = tail;
1131 kunmap_atomic(ring);
1132 flush_dcache_page(ctx->ring_pages[0]);
1133
1134 ctx->completed_events++;
1135 if (ctx->completed_events > 1)
1136 refill_reqs_available(ctx, head, tail);
1137 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1138
1139 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1140
1141 /*
1142 * Check if the user asked us to deliver the result through an
1143 * eventfd. The eventfd_signal() function is safe to be called
1144 * from IRQ context.
1145 */
1146 if (iocb->ki_eventfd)
1147 eventfd_signal(iocb->ki_eventfd, 1);
1148
1149 /*
1150 * We have to order our ring_info tail store above and test
1151 * of the wait list below outside the wait lock. This is
1152 * like in wake_up_bit() where clearing a bit has to be
1153 * ordered with the unlocked test.
1154 */
1155 smp_mb();
1156
1157 if (waitqueue_active(&ctx->wait))
1158 wake_up(&ctx->wait);
1159}
1160
1161static inline void iocb_put(struct aio_kiocb *iocb)
1162{
1163 if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1164 aio_complete(iocb);
1165 iocb_destroy(iocb);
1166 }
1167}
1168
1169/* aio_read_events_ring
1170 * Pull an event off of the ioctx's event ring. Returns the number of
1171 * events fetched
1172 */
1173static long aio_read_events_ring(struct kioctx *ctx,
1174 struct io_event __user *event, long nr)
1175{
1176 struct aio_ring *ring;
1177 unsigned head, tail, pos;
1178 long ret = 0;
1179 int copy_ret;
1180
1181 /*
1182 * The mutex can block and wake us up and that will cause
1183 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1184 * and repeat. This should be rare enough that it doesn't cause
1185 * peformance issues. See the comment in read_events() for more detail.
1186 */
1187 sched_annotate_sleep();
1188 mutex_lock(&ctx->ring_lock);
1189
1190 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1191 ring = kmap_atomic(ctx->ring_pages[0]);
1192 head = ring->head;
1193 tail = ring->tail;
1194 kunmap_atomic(ring);
1195
1196 /*
1197 * Ensure that once we've read the current tail pointer, that
1198 * we also see the events that were stored up to the tail.
1199 */
1200 smp_rmb();
1201
1202 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1203
1204 if (head == tail)
1205 goto out;
1206
1207 head %= ctx->nr_events;
1208 tail %= ctx->nr_events;
1209
1210 while (ret < nr) {
1211 long avail;
1212 struct io_event *ev;
1213 struct page *page;
1214
1215 avail = (head <= tail ? tail : ctx->nr_events) - head;
1216 if (head == tail)
1217 break;
1218
1219 pos = head + AIO_EVENTS_OFFSET;
1220 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1221 pos %= AIO_EVENTS_PER_PAGE;
1222
1223 avail = min(avail, nr - ret);
1224 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1225
1226 ev = kmap(page);
1227 copy_ret = copy_to_user(event + ret, ev + pos,
1228 sizeof(*ev) * avail);
1229 kunmap(page);
1230
1231 if (unlikely(copy_ret)) {
1232 ret = -EFAULT;
1233 goto out;
1234 }
1235
1236 ret += avail;
1237 head += avail;
1238 head %= ctx->nr_events;
1239 }
1240
1241 ring = kmap_atomic(ctx->ring_pages[0]);
1242 ring->head = head;
1243 kunmap_atomic(ring);
1244 flush_dcache_page(ctx->ring_pages[0]);
1245
1246 pr_debug("%li h%u t%u\n", ret, head, tail);
1247out:
1248 mutex_unlock(&ctx->ring_lock);
1249
1250 return ret;
1251}
1252
1253static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1254 struct io_event __user *event, long *i)
1255{
1256 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1257
1258 if (ret > 0)
1259 *i += ret;
1260
1261 if (unlikely(atomic_read(&ctx->dead)))
1262 ret = -EINVAL;
1263
1264 if (!*i)
1265 *i = ret;
1266
1267 return ret < 0 || *i >= min_nr;
1268}
1269
1270static long read_events(struct kioctx *ctx, long min_nr, long nr,
1271 struct io_event __user *event,
1272 ktime_t until)
1273{
1274 long ret = 0;
1275
1276 /*
1277 * Note that aio_read_events() is being called as the conditional - i.e.
1278 * we're calling it after prepare_to_wait() has set task state to
1279 * TASK_INTERRUPTIBLE.
1280 *
1281 * But aio_read_events() can block, and if it blocks it's going to flip
1282 * the task state back to TASK_RUNNING.
1283 *
1284 * This should be ok, provided it doesn't flip the state back to
1285 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1286 * will only happen if the mutex_lock() call blocks, and we then find
1287 * the ringbuffer empty. So in practice we should be ok, but it's
1288 * something to be aware of when touching this code.
1289 */
1290 if (until == 0)
1291 aio_read_events(ctx, min_nr, nr, event, &ret);
1292 else
1293 wait_event_interruptible_hrtimeout(ctx->wait,
1294 aio_read_events(ctx, min_nr, nr, event, &ret),
1295 until);
1296 return ret;
1297}
1298
1299/* sys_io_setup:
1300 * Create an aio_context capable of receiving at least nr_events.
1301 * ctxp must not point to an aio_context that already exists, and
1302 * must be initialized to 0 prior to the call. On successful
1303 * creation of the aio_context, *ctxp is filled in with the resulting
1304 * handle. May fail with -EINVAL if *ctxp is not initialized,
1305 * if the specified nr_events exceeds internal limits. May fail
1306 * with -EAGAIN if the specified nr_events exceeds the user's limit
1307 * of available events. May fail with -ENOMEM if insufficient kernel
1308 * resources are available. May fail with -EFAULT if an invalid
1309 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1310 * implemented.
1311 */
1312SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1313{
1314 struct kioctx *ioctx = NULL;
1315 unsigned long ctx;
1316 long ret;
1317
1318 ret = get_user(ctx, ctxp);
1319 if (unlikely(ret))
1320 goto out;
1321
1322 ret = -EINVAL;
1323 if (unlikely(ctx || nr_events == 0)) {
1324 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1325 ctx, nr_events);
1326 goto out;
1327 }
1328
1329 ioctx = ioctx_alloc(nr_events);
1330 ret = PTR_ERR(ioctx);
1331 if (!IS_ERR(ioctx)) {
1332 ret = put_user(ioctx->user_id, ctxp);
1333 if (ret)
1334 kill_ioctx(current->mm, ioctx, NULL);
1335 percpu_ref_put(&ioctx->users);
1336 }
1337
1338out:
1339 return ret;
1340}
1341
1342#ifdef CONFIG_COMPAT
1343COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1344{
1345 struct kioctx *ioctx = NULL;
1346 unsigned long ctx;
1347 long ret;
1348
1349 ret = get_user(ctx, ctx32p);
1350 if (unlikely(ret))
1351 goto out;
1352
1353 ret = -EINVAL;
1354 if (unlikely(ctx || nr_events == 0)) {
1355 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1356 ctx, nr_events);
1357 goto out;
1358 }
1359
1360 ioctx = ioctx_alloc(nr_events);
1361 ret = PTR_ERR(ioctx);
1362 if (!IS_ERR(ioctx)) {
1363 /* truncating is ok because it's a user address */
1364 ret = put_user((u32)ioctx->user_id, ctx32p);
1365 if (ret)
1366 kill_ioctx(current->mm, ioctx, NULL);
1367 percpu_ref_put(&ioctx->users);
1368 }
1369
1370out:
1371 return ret;
1372}
1373#endif
1374
1375/* sys_io_destroy:
1376 * Destroy the aio_context specified. May cancel any outstanding
1377 * AIOs and block on completion. Will fail with -ENOSYS if not
1378 * implemented. May fail with -EINVAL if the context pointed to
1379 * is invalid.
1380 */
1381SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1382{
1383 struct kioctx *ioctx = lookup_ioctx(ctx);
1384 if (likely(NULL != ioctx)) {
1385 struct ctx_rq_wait wait;
1386 int ret;
1387
1388 init_completion(&wait.comp);
1389 atomic_set(&wait.count, 1);
1390
1391 /* Pass requests_done to kill_ioctx() where it can be set
1392 * in a thread-safe way. If we try to set it here then we have
1393 * a race condition if two io_destroy() called simultaneously.
1394 */
1395 ret = kill_ioctx(current->mm, ioctx, &wait);
1396 percpu_ref_put(&ioctx->users);
1397
1398 /* Wait until all IO for the context are done. Otherwise kernel
1399 * keep using user-space buffers even if user thinks the context
1400 * is destroyed.
1401 */
1402 if (!ret)
1403 wait_for_completion(&wait.comp);
1404
1405 return ret;
1406 }
1407 pr_debug("EINVAL: invalid context id\n");
1408 return -EINVAL;
1409}
1410
1411static void aio_remove_iocb(struct aio_kiocb *iocb)
1412{
1413 struct kioctx *ctx = iocb->ki_ctx;
1414 unsigned long flags;
1415
1416 spin_lock_irqsave(&ctx->ctx_lock, flags);
1417 list_del(&iocb->ki_list);
1418 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1419}
1420
1421static void aio_complete_rw(struct kiocb *kiocb, long res, long res2)
1422{
1423 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1424
1425 if (!list_empty_careful(&iocb->ki_list))
1426 aio_remove_iocb(iocb);
1427
1428 if (kiocb->ki_flags & IOCB_WRITE) {
1429 struct inode *inode = file_inode(kiocb->ki_filp);
1430
1431 /*
1432 * Tell lockdep we inherited freeze protection from submission
1433 * thread.
1434 */
1435 if (S_ISREG(inode->i_mode))
1436 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1437 file_end_write(kiocb->ki_filp);
1438 }
1439
1440 iocb->ki_res.res = res;
1441 iocb->ki_res.res2 = res2;
1442 iocb_put(iocb);
1443}
1444
1445static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
1446{
1447 int ret;
1448
1449 req->ki_complete = aio_complete_rw;
1450 req->private = NULL;
1451 req->ki_pos = iocb->aio_offset;
1452 req->ki_flags = iocb_flags(req->ki_filp);
1453 if (iocb->aio_flags & IOCB_FLAG_RESFD)
1454 req->ki_flags |= IOCB_EVENTFD;
1455 req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp));
1456 if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1457 /*
1458 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1459 * aio_reqprio is interpreted as an I/O scheduling
1460 * class and priority.
1461 */
1462 ret = ioprio_check_cap(iocb->aio_reqprio);
1463 if (ret) {
1464 pr_debug("aio ioprio check cap error: %d\n", ret);
1465 return ret;
1466 }
1467
1468 req->ki_ioprio = iocb->aio_reqprio;
1469 } else
1470 req->ki_ioprio = get_current_ioprio();
1471
1472 ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1473 if (unlikely(ret))
1474 return ret;
1475
1476 req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1477 return 0;
1478}
1479
1480static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
1481 struct iovec **iovec, bool vectored, bool compat,
1482 struct iov_iter *iter)
1483{
1484 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1485 size_t len = iocb->aio_nbytes;
1486
1487 if (!vectored) {
1488 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1489 *iovec = NULL;
1490 return ret;
1491 }
1492#ifdef CONFIG_COMPAT
1493 if (compat)
1494 return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec,
1495 iter);
1496#endif
1497 return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter);
1498}
1499
1500static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1501{
1502 switch (ret) {
1503 case -EIOCBQUEUED:
1504 break;
1505 case -ERESTARTSYS:
1506 case -ERESTARTNOINTR:
1507 case -ERESTARTNOHAND:
1508 case -ERESTART_RESTARTBLOCK:
1509 /*
1510 * There's no easy way to restart the syscall since other AIO's
1511 * may be already running. Just fail this IO with EINTR.
1512 */
1513 ret = -EINTR;
1514 /*FALLTHRU*/
1515 default:
1516 req->ki_complete(req, ret, 0);
1517 }
1518}
1519
1520static int aio_read(struct kiocb *req, const struct iocb *iocb,
1521 bool vectored, bool compat)
1522{
1523 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1524 struct iov_iter iter;
1525 struct file *file;
1526 int ret;
1527
1528 ret = aio_prep_rw(req, iocb);
1529 if (ret)
1530 return ret;
1531 file = req->ki_filp;
1532 if (unlikely(!(file->f_mode & FMODE_READ)))
1533 return -EBADF;
1534 ret = -EINVAL;
1535 if (unlikely(!file->f_op->read_iter))
1536 return -EINVAL;
1537
1538 ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1539 if (ret < 0)
1540 return ret;
1541 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1542 if (!ret)
1543 aio_rw_done(req, call_read_iter(file, req, &iter));
1544 kfree(iovec);
1545 return ret;
1546}
1547
1548static int aio_write(struct kiocb *req, const struct iocb *iocb,
1549 bool vectored, bool compat)
1550{
1551 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1552 struct iov_iter iter;
1553 struct file *file;
1554 int ret;
1555
1556 ret = aio_prep_rw(req, iocb);
1557 if (ret)
1558 return ret;
1559 file = req->ki_filp;
1560
1561 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1562 return -EBADF;
1563 if (unlikely(!file->f_op->write_iter))
1564 return -EINVAL;
1565
1566 ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1567 if (ret < 0)
1568 return ret;
1569 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1570 if (!ret) {
1571 /*
1572 * Open-code file_start_write here to grab freeze protection,
1573 * which will be released by another thread in
1574 * aio_complete_rw(). Fool lockdep by telling it the lock got
1575 * released so that it doesn't complain about the held lock when
1576 * we return to userspace.
1577 */
1578 if (S_ISREG(file_inode(file)->i_mode)) {
1579 __sb_start_write(file_inode(file)->i_sb, SB_FREEZE_WRITE, true);
1580 __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1581 }
1582 req->ki_flags |= IOCB_WRITE;
1583 aio_rw_done(req, call_write_iter(file, req, &iter));
1584 }
1585 kfree(iovec);
1586 return ret;
1587}
1588
1589static void aio_fsync_work(struct work_struct *work)
1590{
1591 struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1592
1593 iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1594 iocb_put(iocb);
1595}
1596
1597static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1598 bool datasync)
1599{
1600 if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1601 iocb->aio_rw_flags))
1602 return -EINVAL;
1603
1604 if (unlikely(!req->file->f_op->fsync))
1605 return -EINVAL;
1606
1607 req->datasync = datasync;
1608 INIT_WORK(&req->work, aio_fsync_work);
1609 schedule_work(&req->work);
1610 return 0;
1611}
1612
1613static void aio_poll_complete_work(struct work_struct *work)
1614{
1615 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1616 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1617 struct poll_table_struct pt = { ._key = req->events };
1618 struct kioctx *ctx = iocb->ki_ctx;
1619 __poll_t mask = 0;
1620
1621 if (!READ_ONCE(req->cancelled))
1622 mask = vfs_poll(req->file, &pt) & req->events;
1623
1624 /*
1625 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1626 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1627 * synchronize with them. In the cancellation case the list_del_init
1628 * itself is not actually needed, but harmless so we keep it in to
1629 * avoid further branches in the fast path.
1630 */
1631 spin_lock_irq(&ctx->ctx_lock);
1632 if (!mask && !READ_ONCE(req->cancelled)) {
1633 add_wait_queue(req->head, &req->wait);
1634 spin_unlock_irq(&ctx->ctx_lock);
1635 return;
1636 }
1637 list_del_init(&iocb->ki_list);
1638 iocb->ki_res.res = mangle_poll(mask);
1639 req->done = true;
1640 spin_unlock_irq(&ctx->ctx_lock);
1641
1642 iocb_put(iocb);
1643}
1644
1645/* assumes we are called with irqs disabled */
1646static int aio_poll_cancel(struct kiocb *iocb)
1647{
1648 struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1649 struct poll_iocb *req = &aiocb->poll;
1650
1651 spin_lock(&req->head->lock);
1652 WRITE_ONCE(req->cancelled, true);
1653 if (!list_empty(&req->wait.entry)) {
1654 list_del_init(&req->wait.entry);
1655 schedule_work(&aiocb->poll.work);
1656 }
1657 spin_unlock(&req->head->lock);
1658
1659 return 0;
1660}
1661
1662static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1663 void *key)
1664{
1665 struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1666 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1667 __poll_t mask = key_to_poll(key);
1668 unsigned long flags;
1669
1670 /* for instances that support it check for an event match first: */
1671 if (mask && !(mask & req->events))
1672 return 0;
1673
1674 list_del_init(&req->wait.entry);
1675
1676 if (mask && spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1677 /*
1678 * Try to complete the iocb inline if we can. Use
1679 * irqsave/irqrestore because not all filesystems (e.g. fuse)
1680 * call this function with IRQs disabled and because IRQs
1681 * have to be disabled before ctx_lock is obtained.
1682 */
1683 list_del(&iocb->ki_list);
1684 iocb->ki_res.res = mangle_poll(mask);
1685 req->done = true;
1686 spin_unlock_irqrestore(&iocb->ki_ctx->ctx_lock, flags);
1687 iocb_put(iocb);
1688 } else {
1689 schedule_work(&req->work);
1690 }
1691 return 1;
1692}
1693
1694struct aio_poll_table {
1695 struct poll_table_struct pt;
1696 struct aio_kiocb *iocb;
1697 int error;
1698};
1699
1700static void
1701aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1702 struct poll_table_struct *p)
1703{
1704 struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1705
1706 /* multiple wait queues per file are not supported */
1707 if (unlikely(pt->iocb->poll.head)) {
1708 pt->error = -EINVAL;
1709 return;
1710 }
1711
1712 pt->error = 0;
1713 pt->iocb->poll.head = head;
1714 add_wait_queue(head, &pt->iocb->poll.wait);
1715}
1716
1717static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1718{
1719 struct kioctx *ctx = aiocb->ki_ctx;
1720 struct poll_iocb *req = &aiocb->poll;
1721 struct aio_poll_table apt;
1722 bool cancel = false;
1723 __poll_t mask;
1724
1725 /* reject any unknown events outside the normal event mask. */
1726 if ((u16)iocb->aio_buf != iocb->aio_buf)
1727 return -EINVAL;
1728 /* reject fields that are not defined for poll */
1729 if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1730 return -EINVAL;
1731
1732 INIT_WORK(&req->work, aio_poll_complete_work);
1733 req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1734
1735 req->head = NULL;
1736 req->done = false;
1737 req->cancelled = false;
1738
1739 apt.pt._qproc = aio_poll_queue_proc;
1740 apt.pt._key = req->events;
1741 apt.iocb = aiocb;
1742 apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1743
1744 /* initialized the list so that we can do list_empty checks */
1745 INIT_LIST_HEAD(&req->wait.entry);
1746 init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1747
1748 mask = vfs_poll(req->file, &apt.pt) & req->events;
1749 spin_lock_irq(&ctx->ctx_lock);
1750 if (likely(req->head)) {
1751 spin_lock(&req->head->lock);
1752 if (unlikely(list_empty(&req->wait.entry))) {
1753 if (apt.error)
1754 cancel = true;
1755 apt.error = 0;
1756 mask = 0;
1757 }
1758 if (mask || apt.error) {
1759 list_del_init(&req->wait.entry);
1760 } else if (cancel) {
1761 WRITE_ONCE(req->cancelled, true);
1762 } else if (!req->done) { /* actually waiting for an event */
1763 list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1764 aiocb->ki_cancel = aio_poll_cancel;
1765 }
1766 spin_unlock(&req->head->lock);
1767 }
1768 if (mask) { /* no async, we'd stolen it */
1769 aiocb->ki_res.res = mangle_poll(mask);
1770 apt.error = 0;
1771 }
1772 spin_unlock_irq(&ctx->ctx_lock);
1773 if (mask)
1774 iocb_put(aiocb);
1775 return apt.error;
1776}
1777
1778static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1779 struct iocb __user *user_iocb, struct aio_kiocb *req,
1780 bool compat)
1781{
1782 req->ki_filp = fget(iocb->aio_fildes);
1783 if (unlikely(!req->ki_filp))
1784 return -EBADF;
1785
1786 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1787 struct eventfd_ctx *eventfd;
1788 /*
1789 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1790 * instance of the file* now. The file descriptor must be
1791 * an eventfd() fd, and will be signaled for each completed
1792 * event using the eventfd_signal() function.
1793 */
1794 eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
1795 if (IS_ERR(eventfd))
1796 return PTR_ERR(eventfd);
1797
1798 req->ki_eventfd = eventfd;
1799 }
1800
1801 if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
1802 pr_debug("EFAULT: aio_key\n");
1803 return -EFAULT;
1804 }
1805
1806 req->ki_res.obj = (u64)(unsigned long)user_iocb;
1807 req->ki_res.data = iocb->aio_data;
1808 req->ki_res.res = 0;
1809 req->ki_res.res2 = 0;
1810
1811 switch (iocb->aio_lio_opcode) {
1812 case IOCB_CMD_PREAD:
1813 return aio_read(&req->rw, iocb, false, compat);
1814 case IOCB_CMD_PWRITE:
1815 return aio_write(&req->rw, iocb, false, compat);
1816 case IOCB_CMD_PREADV:
1817 return aio_read(&req->rw, iocb, true, compat);
1818 case IOCB_CMD_PWRITEV:
1819 return aio_write(&req->rw, iocb, true, compat);
1820 case IOCB_CMD_FSYNC:
1821 return aio_fsync(&req->fsync, iocb, false);
1822 case IOCB_CMD_FDSYNC:
1823 return aio_fsync(&req->fsync, iocb, true);
1824 case IOCB_CMD_POLL:
1825 return aio_poll(req, iocb);
1826 default:
1827 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1828 return -EINVAL;
1829 }
1830}
1831
1832static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1833 bool compat)
1834{
1835 struct aio_kiocb *req;
1836 struct iocb iocb;
1837 int err;
1838
1839 if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
1840 return -EFAULT;
1841
1842 /* enforce forwards compatibility on users */
1843 if (unlikely(iocb.aio_reserved2)) {
1844 pr_debug("EINVAL: reserve field set\n");
1845 return -EINVAL;
1846 }
1847
1848 /* prevent overflows */
1849 if (unlikely(
1850 (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
1851 (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
1852 ((ssize_t)iocb.aio_nbytes < 0)
1853 )) {
1854 pr_debug("EINVAL: overflow check\n");
1855 return -EINVAL;
1856 }
1857
1858 req = aio_get_req(ctx);
1859 if (unlikely(!req))
1860 return -EAGAIN;
1861
1862 err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
1863
1864 /* Done with the synchronous reference */
1865 iocb_put(req);
1866
1867 /*
1868 * If err is 0, we'd either done aio_complete() ourselves or have
1869 * arranged for that to be done asynchronously. Anything non-zero
1870 * means that we need to destroy req ourselves.
1871 */
1872 if (unlikely(err)) {
1873 iocb_destroy(req);
1874 put_reqs_available(ctx, 1);
1875 }
1876 return err;
1877}
1878
1879/* sys_io_submit:
1880 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1881 * the number of iocbs queued. May return -EINVAL if the aio_context
1882 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1883 * *iocbpp[0] is not properly initialized, if the operation specified
1884 * is invalid for the file descriptor in the iocb. May fail with
1885 * -EFAULT if any of the data structures point to invalid data. May
1886 * fail with -EBADF if the file descriptor specified in the first
1887 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1888 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1889 * fail with -ENOSYS if not implemented.
1890 */
1891SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1892 struct iocb __user * __user *, iocbpp)
1893{
1894 struct kioctx *ctx;
1895 long ret = 0;
1896 int i = 0;
1897 struct blk_plug plug;
1898
1899 if (unlikely(nr < 0))
1900 return -EINVAL;
1901
1902 ctx = lookup_ioctx(ctx_id);
1903 if (unlikely(!ctx)) {
1904 pr_debug("EINVAL: invalid context id\n");
1905 return -EINVAL;
1906 }
1907
1908 if (nr > ctx->nr_events)
1909 nr = ctx->nr_events;
1910
1911 if (nr > AIO_PLUG_THRESHOLD)
1912 blk_start_plug(&plug);
1913 for (i = 0; i < nr; i++) {
1914 struct iocb __user *user_iocb;
1915
1916 if (unlikely(get_user(user_iocb, iocbpp + i))) {
1917 ret = -EFAULT;
1918 break;
1919 }
1920
1921 ret = io_submit_one(ctx, user_iocb, false);
1922 if (ret)
1923 break;
1924 }
1925 if (nr > AIO_PLUG_THRESHOLD)
1926 blk_finish_plug(&plug);
1927
1928 percpu_ref_put(&ctx->users);
1929 return i ? i : ret;
1930}
1931
1932#ifdef CONFIG_COMPAT
1933COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
1934 int, nr, compat_uptr_t __user *, iocbpp)
1935{
1936 struct kioctx *ctx;
1937 long ret = 0;
1938 int i = 0;
1939 struct blk_plug plug;
1940
1941 if (unlikely(nr < 0))
1942 return -EINVAL;
1943
1944 ctx = lookup_ioctx(ctx_id);
1945 if (unlikely(!ctx)) {
1946 pr_debug("EINVAL: invalid context id\n");
1947 return -EINVAL;
1948 }
1949
1950 if (nr > ctx->nr_events)
1951 nr = ctx->nr_events;
1952
1953 if (nr > AIO_PLUG_THRESHOLD)
1954 blk_start_plug(&plug);
1955 for (i = 0; i < nr; i++) {
1956 compat_uptr_t user_iocb;
1957
1958 if (unlikely(get_user(user_iocb, iocbpp + i))) {
1959 ret = -EFAULT;
1960 break;
1961 }
1962
1963 ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
1964 if (ret)
1965 break;
1966 }
1967 if (nr > AIO_PLUG_THRESHOLD)
1968 blk_finish_plug(&plug);
1969
1970 percpu_ref_put(&ctx->users);
1971 return i ? i : ret;
1972}
1973#endif
1974
1975/* sys_io_cancel:
1976 * Attempts to cancel an iocb previously passed to io_submit. If
1977 * the operation is successfully cancelled, the resulting event is
1978 * copied into the memory pointed to by result without being placed
1979 * into the completion queue and 0 is returned. May fail with
1980 * -EFAULT if any of the data structures pointed to are invalid.
1981 * May fail with -EINVAL if aio_context specified by ctx_id is
1982 * invalid. May fail with -EAGAIN if the iocb specified was not
1983 * cancelled. Will fail with -ENOSYS if not implemented.
1984 */
1985SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1986 struct io_event __user *, result)
1987{
1988 struct kioctx *ctx;
1989 struct aio_kiocb *kiocb;
1990 int ret = -EINVAL;
1991 u32 key;
1992 u64 obj = (u64)(unsigned long)iocb;
1993
1994 if (unlikely(get_user(key, &iocb->aio_key)))
1995 return -EFAULT;
1996 if (unlikely(key != KIOCB_KEY))
1997 return -EINVAL;
1998
1999 ctx = lookup_ioctx(ctx_id);
2000 if (unlikely(!ctx))
2001 return -EINVAL;
2002
2003 spin_lock_irq(&ctx->ctx_lock);
2004 /* TODO: use a hash or array, this sucks. */
2005 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2006 if (kiocb->ki_res.obj == obj) {
2007 ret = kiocb->ki_cancel(&kiocb->rw);
2008 list_del_init(&kiocb->ki_list);
2009 break;
2010 }
2011 }
2012 spin_unlock_irq(&ctx->ctx_lock);
2013
2014 if (!ret) {
2015 /*
2016 * The result argument is no longer used - the io_event is
2017 * always delivered via the ring buffer. -EINPROGRESS indicates
2018 * cancellation is progress:
2019 */
2020 ret = -EINPROGRESS;
2021 }
2022
2023 percpu_ref_put(&ctx->users);
2024
2025 return ret;
2026}
2027
2028static long do_io_getevents(aio_context_t ctx_id,
2029 long min_nr,
2030 long nr,
2031 struct io_event __user *events,
2032 struct timespec64 *ts)
2033{
2034 ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2035 struct kioctx *ioctx = lookup_ioctx(ctx_id);
2036 long ret = -EINVAL;
2037
2038 if (likely(ioctx)) {
2039 if (likely(min_nr <= nr && min_nr >= 0))
2040 ret = read_events(ioctx, min_nr, nr, events, until);
2041 percpu_ref_put(&ioctx->users);
2042 }
2043
2044 return ret;
2045}
2046
2047/* io_getevents:
2048 * Attempts to read at least min_nr events and up to nr events from
2049 * the completion queue for the aio_context specified by ctx_id. If
2050 * it succeeds, the number of read events is returned. May fail with
2051 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2052 * out of range, if timeout is out of range. May fail with -EFAULT
2053 * if any of the memory specified is invalid. May return 0 or
2054 * < min_nr if the timeout specified by timeout has elapsed
2055 * before sufficient events are available, where timeout == NULL
2056 * specifies an infinite timeout. Note that the timeout pointed to by
2057 * timeout is relative. Will fail with -ENOSYS if not implemented.
2058 */
2059#if !defined(CONFIG_64BIT_TIME) || defined(CONFIG_64BIT)
2060
2061SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2062 long, min_nr,
2063 long, nr,
2064 struct io_event __user *, events,
2065 struct __kernel_timespec __user *, timeout)
2066{
2067 struct timespec64 ts;
2068 int ret;
2069
2070 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2071 return -EFAULT;
2072
2073 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2074 if (!ret && signal_pending(current))
2075 ret = -EINTR;
2076 return ret;
2077}
2078
2079#endif
2080
2081struct __aio_sigset {
2082 const sigset_t __user *sigmask;
2083 size_t sigsetsize;
2084};
2085
2086SYSCALL_DEFINE6(io_pgetevents,
2087 aio_context_t, ctx_id,
2088 long, min_nr,
2089 long, nr,
2090 struct io_event __user *, events,
2091 struct __kernel_timespec __user *, timeout,
2092 const struct __aio_sigset __user *, usig)
2093{
2094 struct __aio_sigset ksig = { NULL, };
2095 struct timespec64 ts;
2096 bool interrupted;
2097 int ret;
2098
2099 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2100 return -EFAULT;
2101
2102 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2103 return -EFAULT;
2104
2105 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2106 if (ret)
2107 return ret;
2108
2109 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2110
2111 interrupted = signal_pending(current);
2112 restore_saved_sigmask_unless(interrupted);
2113 if (interrupted && !ret)
2114 ret = -ERESTARTNOHAND;
2115
2116 return ret;
2117}
2118
2119#if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2120
2121SYSCALL_DEFINE6(io_pgetevents_time32,
2122 aio_context_t, ctx_id,
2123 long, min_nr,
2124 long, nr,
2125 struct io_event __user *, events,
2126 struct old_timespec32 __user *, timeout,
2127 const struct __aio_sigset __user *, usig)
2128{
2129 struct __aio_sigset ksig = { NULL, };
2130 struct timespec64 ts;
2131 bool interrupted;
2132 int ret;
2133
2134 if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2135 return -EFAULT;
2136
2137 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2138 return -EFAULT;
2139
2140
2141 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2142 if (ret)
2143 return ret;
2144
2145 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2146
2147 interrupted = signal_pending(current);
2148 restore_saved_sigmask_unless(interrupted);
2149 if (interrupted && !ret)
2150 ret = -ERESTARTNOHAND;
2151
2152 return ret;
2153}
2154
2155#endif
2156
2157#if defined(CONFIG_COMPAT_32BIT_TIME)
2158
2159SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2160 __s32, min_nr,
2161 __s32, nr,
2162 struct io_event __user *, events,
2163 struct old_timespec32 __user *, timeout)
2164{
2165 struct timespec64 t;
2166 int ret;
2167
2168 if (timeout && get_old_timespec32(&t, timeout))
2169 return -EFAULT;
2170
2171 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2172 if (!ret && signal_pending(current))
2173 ret = -EINTR;
2174 return ret;
2175}
2176
2177#endif
2178
2179#ifdef CONFIG_COMPAT
2180
2181struct __compat_aio_sigset {
2182 compat_uptr_t sigmask;
2183 compat_size_t sigsetsize;
2184};
2185
2186#if defined(CONFIG_COMPAT_32BIT_TIME)
2187
2188COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2189 compat_aio_context_t, ctx_id,
2190 compat_long_t, min_nr,
2191 compat_long_t, nr,
2192 struct io_event __user *, events,
2193 struct old_timespec32 __user *, timeout,
2194 const struct __compat_aio_sigset __user *, usig)
2195{
2196 struct __compat_aio_sigset ksig = { 0, };
2197 struct timespec64 t;
2198 bool interrupted;
2199 int ret;
2200
2201 if (timeout && get_old_timespec32(&t, timeout))
2202 return -EFAULT;
2203
2204 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2205 return -EFAULT;
2206
2207 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2208 if (ret)
2209 return ret;
2210
2211 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2212
2213 interrupted = signal_pending(current);
2214 restore_saved_sigmask_unless(interrupted);
2215 if (interrupted && !ret)
2216 ret = -ERESTARTNOHAND;
2217
2218 return ret;
2219}
2220
2221#endif
2222
2223COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2224 compat_aio_context_t, ctx_id,
2225 compat_long_t, min_nr,
2226 compat_long_t, nr,
2227 struct io_event __user *, events,
2228 struct __kernel_timespec __user *, timeout,
2229 const struct __compat_aio_sigset __user *, usig)
2230{
2231 struct __compat_aio_sigset ksig = { 0, };
2232 struct timespec64 t;
2233 bool interrupted;
2234 int ret;
2235
2236 if (timeout && get_timespec64(&t, timeout))
2237 return -EFAULT;
2238
2239 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2240 return -EFAULT;
2241
2242 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2243 if (ret)
2244 return ret;
2245
2246 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2247
2248 interrupted = signal_pending(current);
2249 restore_saved_sigmask_unless(interrupted);
2250 if (interrupted && !ret)
2251 ret = -ERESTARTNOHAND;
2252
2253 return ret;
2254}
2255#endif